Skip to content

CHANGELOG.md

Latest commit

 

History

History
752 lines (613 loc) · 80.1 KB

CHANGELOG.md

File metadata and controls

752 lines (613 loc) · 80.1 KB
Raw

Changelog

This is a synthesized, agent-facing changelog for the full history of franken_engine.

Scope window: project inception on 2026-02-18 through HEAD cf700313c (2026-08-31). The dated post-snapshot sections below record the post-b1f5bc91c work (lockstep differential, evidence freshness, lint debt); the original “Current window (2026-07-26 → 2026-08-19)” section above is kept verbatim as the historical record at that point. Base synthesis covered inception through d51f2715 (2026-05-15). Dated post-snapshot sections below carry the later release history forward, including the current window after 7ec31d156 (2026-07-25).

The base synthesis was rebuilt from git history (4,446 commits and no published releases at that snapshot), the checked-in beads tracker (.beads/issues.jsonl), the in-tree docs/CLAIM_TO_PROOF_MATRIX_V1.md claim ledger, and the contemporaneous workstream notes left in docs/ and runbooks/; the dated sections below carry the later release history forward.

The first conventional release, v0.1.0, was published on 2026-05-29. Current main also uses artifact-bundle versions for individual decision, benchmark, and replay claims; those schema versions remain independent of Cargo package semver.

The earlier current-window catch-up below covers 7ec31d156 (2026-07-25) through b1f5bc91c (2026-08-19): 513 non-merge commits. There is still only one GitHub Release (v0.1.0). The post-b1f5bc91c work documented in the dated section below adds the lockstep comparator fix, evidence-freshness restoration, seven clippy lints closed, three dead CLI placeholders wired, and two product mock seams evicted. Cargo frankenengine-engine / frankenengine-core remain on the unreleased 0.2.0 compatibility-staging line; that number is not a tag and not a Release.

Post-Snapshot Update — Current window (2026-07-26 → 2026-08-19)

After the sibling-gating correction of 2026-07-25, the next month is the product-authority and evidence-honesty wave: host effects stop being ambient, generated code gets its own realm, IFC labels survive the heap, and the engine finally consumes published sibling crates instead of /dp path deps. HEAD is b1f5bc91c (2026-08-19).

Delivered capability

  • Checkout-independent sibling isolation (bd-gw4cg). The nine FrankenSuite siblings that used to be /dp path deps now resolve from crates.io. CI gained a standalone-release lane that must compile without a sibling checkout. This is the follow-through on the 2026-07-25 claim correction: --no-default-features still requires manifests to exist, but those manifests are now registry crates, not local trees.
  • Evidence freshness and split signing authority. Claim freshness became risk-weighted (drift signals, tiers, a schedule) and fail-closed at release. Live orchestrator batches seal as chained artifacts. Runtime vs Lab signing authority split; the ledger schema moved to v2. Generated Function objects got a realm-owned store with content-addressed provenance and isolated globals (bd-fw7zd.8).
  • Execution cells, adaptive routing, timers. A sealed Tier-I cell is now the authority boundary for interpreter instruction use, host I/O, and process spawn. Adaptive routing v2 binds profile selection before dispatch, with a SafeMode and a state budget (bd-wftp1). Timers run through two-phase cell authority and refuse until an interpreter event-loop is bound.
  • Module load, EventEmitter identity, RandomRead entropy. ImportModule and require share one module_load capability key (bd-iyp3h). EventEmitter is a constructible identity with authenticated instanceof so cluster identity works (bd-dspwz). Cryptographic entropy is its own RandomRead grant, lowered through typed hostcalls (bd-opsnv). Process spawn gained sealed verified-file identity (bd-m5p8q) and stable NotFound discriminators (bd-x85a7.3). require('url') helpers lower behind a usage gate (bd-4awsz).
  • Prepared Hybrid eval and interpreter hot-path cuts. Hybrid eval compiles once and reuses immutable IR3. Ordinary own-property reads skip proxy/IFC scans; dense array growth/pop skip length scans; ephemeral traces hand off without cloning payloads. Differential-oracle reports are v2 with honest lifecycles.
  • Janitor docs-reorg (2026-08-19). Root reports moved under docs/planning/; skill-loop scratch is untracked. See the janitor wave below.

Closed workstreams (selected)

bd-gw4cg, bd-fw7zd.8, bd-wftp1, bd-iyp3h, bd-dspwz, bd-opsnv, bd-m5p8q, bd-x85a7.1 / .3, bd-padqo, bd-ojvo1, bd-4awsz.

Representative commits

  • 5a40110e5 — build(deps): resolve the nine sibling crates from crates.io, not /dp paths (bd-gw4cg)
  • 3526cf9d4 — ci: enforce checkout-independent standalone release
  • 61d4d7591 — feat(evidence): make freshness actually risk-weighted -- drift signals, tiers, schedule (BRIDGE-19.18)
  • dedd88284 — feat(evidence): split Runtime vs Lab signing authority; ledger schema v2
  • ad2d6f176 — feat(engine): isolate generated-function realm globals and rotate identities
  • 57dbc1d08 — feat(cell): seal a Tier-I interpreter and effect authority boundary
  • fd7c61e5e — feat(engine): harden adaptive routing v2 — budgeted state, staged learn (bd-wftp1)
  • 4c06e8402 — feat(timer): run interpreter timers through a two-phase cell authority
  • 0e4e1c20e — feat(capability): gate ImportModule and require on one module_load key (bd-iyp3h)
  • 75cf4822d — feat(events): give EventEmitter a constructible identity and authenticated instanceof (bd-dspwz)
  • 050b77e2c — feat(crypto): lower authenticated entropy through typed RandomRead hostcalls (bd-opsnv)
  • 9781a7dd1 — feat(eval): compile Hybrid eval once and reuse immutable IR3
  • 17c7fb24c — feat(url): lower usage-gated require('url') helpers (bd-4awsz)

Post-Snapshot Update — Lockstep Differential, Evidence Freshness, and Lint-Debt Catch-up (2026-08-20 → 2026-08-31)

A focused catch-up window. The “Current window” section above stops at 2026-08-19; the dated entries below it (Janitor docs-reorg) cover the 2026-08-19 cleanup. The work in this new section runs from 2026-08-20 through HEAD cf700313c (2026-08-31). It is not a full multi-agent synthesis of the window — that belongs in a separate Standard-Rebuild pass — but documents the work this reality check drove end-to-end and the real findings it surfaced. For other workstreams in the window (TLS bd-il0d9 family, IFC capture-origins

  • Theme 6 — lockstep comparator + parser fix verified at a694fba6d. The new post-b1f5bc91c window in this same section also drove (a) the multi-engine harness redesign so the lockstep differential reports real signal instead of a vacuous 8/8 spurious, and (b) the parser fix that closed the one real drift (script_await, top-level await in script goal). With both fixes in, the default 8-fixture lockstep reports 8/8 equivalent, 0 divergences (verified 2026-08-31 in a clean worktree at a694fba6d; binary franken_lockstep_runner built release with the pinned nightly-2026-08-25 toolchain; fixture set crates/franken-engine/tests/fixtures/parser_phase0_semantic_fixtures.json).
  • Lockstep comparator made honest (bd-lockstep-canonical-ast-divergence-zq6lo, resolved). The frankenctl test lockstep command, the frankenctl oracle differential, and the parser_multi_engine_harness all share one equivalence comparator. Before this work the comparator was blind: the Node/Bun adapter scripts emitted sha256(normalize-newlines(source)) and franken_canonical emitted a canonical-AST digest, and the signature grouped them by hash, so the default 8 fixtures reported 8/8 spurious critical divergences with zero real signal. The fix: adapters now do compile-only vm.Script / vm.SourceTextModule syntax validation and emit a real parse verdict; the comparator groups on parse-verdict + diagnostic digest; AST digests are compared only within their own AstSpace (internal = golden ↔ franken; external = source fingerprint). Once the parser fix at a694fba6d landed, the default 8-fixture lockstep returns to 8/8 equivalent, 0 divergences (verified 2026-08-31 — see Theme 6).
  • Seven pre-existing clippy -D warnings failures closed (5cdea2e30). The AGENTS clippy gate was red repo-wide since 7291545a5 (2026-07-20): one needless_bool in franken-core/src/ts_normalization.rs, plus five in the engine lib (field_reassign_with_default, type_complexity + new alias, needless_range_loop, manual_is_multiple_of, chunks_exact→as_chunks, plus another needless_bool), plus the dead CODE_UNSUPPORTED_PLACEHOLDER_COMMAND constant orphaned by the placeholder-wiring work. All fixed with the clippy-suggested forms; no behavior change. cargo clippy --release -p frankenengine-engine --lib --bin frankenctl --bin franken_lockstep_runner -- -D warnings was green at the time of these fixes; the gate remains red today only because of an in-progress unclosed-delimiter edit in crates/franken-engine/src/lowering_pipeline.rs:43553 from another agent — that is not a regression in any committed change.
  • All 13 OBSERVED claim receipts refreshed at HEAD (9174376c0). After the work in the prior tranche, the receipt-freshness check went from fresh=0/13 (the original reality-check finding) to fresh=13/13, stale=0, drift clean. All three tiers re-verified end-to-end via run_evidence_refresh_schedule.sh: volatile 6/6 (FE-CLAIM-001, 006, 007, 015, 022, 024), standard 5/5 (002, 003, 004, 013, 025), frozen 2/2 (008, 009). Each receipt's repro.lock partner is present; the claim_to_proof_matrix_gate.sh ci ci-mode check is clean.
  • Three dead CLI placeholders wired to real implementations (365e430fd). frankenctl reports lowering-gap now calls lowering_gap_inventory::write_lowering_gap_inventory_bundle in-process (full replay-shaped bundle, 7 sites, deterministic hash). frankenctl gates signature-drift delegates to the canonical franken_signature_drift_gate binary via a new exe-relative/PATH sibling resolver; --config is explicitly refused (the canonical gate has no config input). frankenctl test lockstep delegates to the canonical franken_lockstep_runner; --config maps to --runtime-specs. The old fail_closed_placeholder_command + its error code constant were removed (clean cutover).
  • Two product-code mock seams evicted. HostcallMigrationAdapter::new() default stack is now console-only; fs:read/fs:write effects fail closed (capability-gated) until a handler is installed; tests install MockFsHandler explicitly with a new fail-closed regression test. get_platform_environment no longer fabricates "test" / "1.75.0"; host platform is really captured via a new LinuxX64EnvCapture (plus shared rustc-verbose parser deduped across all three platform captures), and remote platforms carry explicit "unavailable" markers instead of plausible fakes.
  • Placeholder-closure contract references repaired. Both copies of docs/rgc_placeholder_closure_verification_v1.json and the MD companion previously referenced four run_placeholder_*.sh scripts that did not exist; now they point to the real zero-placeholder gate surfaces (run_rgc_zero_placeholder_gate.sh ci and frankenctl gates zero-placeholder).

Real findings the work surfaced

  • bd-franken-parser-top-level-await-script-sapi0 (P1, parser team — fixed by a694fba6d "feat(parser): enforce ES2020 await context rules across goals and async functions"). The first genuine drift the lockstep differential surfaced: fixture script_await — franken_canonical accepted top-level await in a classic script goal (parse: ok), while Node and Bun correctly rejected it (parse: syntax_error). Per ECMA-262, top-level await is only valid at the top level of a module (sourceType: "module"); in script it is a SyntaxError. The bead records the reproduction command + per-engine parse verdicts for posterity; with the corrected parser the default 8-fixture lockstep should return to 8/8 equivalent once re-verified.
  • Pre-existing test failure recorded (bd-circular-dep-test-preexisting-5x3hy, P3). algebraic_effects_integration::test_circular_dependency_detection asserts an impossible CircularDependency from an empty HandlerStack (its own TODO admits the mechanism became untestable when dependency_path went private). Verified present and unchanged at parent 08aa44884; filed per the repo's pre-existing-failure convention (bd-ggxm8 / bd-y9p6y / bd-3ltox / bd-xulus pattern). Not fixed here; correct fix requires either a test-visible constructor for dependency_path or moving the test into the lib's #[cfg(test)] where the private field is reachable.
  • Toolchain env-blocker documented. The pinned nightly-2026-08-25 + PATH-pinned toolchain combination is what makes cargo clippy and cargo build work cleanly in this environment; without PATH pinning, cargo's clippy invocation pulls the floating-nightly clippy-driver which carries a metadata-version skew and produces false E0514 cross-version artifacts. The rch fleet preflight refuses all workers (hard_preflight=12) despite rch doctor reporting 32/32 checks pass — an operator-side disagreement to triage. rustup component add i686-unknown-linux-linux-musl is currently broken on the shared stable channel (Scrt1.o conflict) — unrelated to this repo; operator fix is rustup toolchain uninstall stable && reinstall or manual removal of the stale files.

Closed workstreams (selected)

bd-lockstep-canonical-ast-divergence-zq6lo (comparator artifact root-caused and fixed; the residual 1/8 divergence is the script_await parser question above, filed separately). bd-cli-wiring-mock-seam-honesty-d3e9s (full evidence trail in the bead close reason; 5/6 command-line surfaces wired, 6 mock/contract fixes, all verified at commit 365e430fd). bd-three-feature-floor-owner-2fjbh (the only remaining unowned vision item after BRIDGE-02.5 turned out to already own the weighted-denominator rerun). bd-franken-parser-top-level-await-script-sapi0 (script_await parser over-permit fixed by a694fba6d; re-verification pending the lower_pipeline WIP build to clear).

Representative commits

  • 365e430fd — feat(engine,cli): replace operator placeholders with sibling resolution and fail-closed env capture
  • 5cdea2e30 — fix(lints): resolve 7 clippy -D warnings failures blocking the AGENTS clippy gate
  • 63d8b7594 — feat(lockstep,harness): add compile-only syntax verdicts to Node/Bun adapters and multi-engine harness
  • f876ef1c8 — test(lockstep,harness): update multi-engine harness and lockstep integration tests for parse verdicts
  • 9174376c0 — chore(evidence): refresh all 13 OBSERVED claim receipts at current HEAD

Note. The remaining lowering_pipeline.rs:43553 unclosed delimiter is from another agent's uncommitted in-progress edit. The committed state at 80c26bf0f builds clean (cargo check --lib exit 0); only NEW test invocations that re-link the lib are affected. Not a regression of any change documented here.

Post-Snapshot Update — Team work in the same window (2026-08-20 → 2026-08-31)

A separate, agent-recorded catch-up of the work other teams drove in the same window. This is a research-led summary, not an independently-verified synthesis — bead close-reasons and commit messages were sampled for thematic clustering; specific behavior claims are owned by the originating beads. For a full Standard-Rebuild pass with link verification and a coverage ledger, this section is the index, not the audit.

The wave that began with the 2026-08-19 product-authority push (charter §2, charter §7) closed into five cross-cutting themes during this window: (1) live-Node / live-bun hardening of TLS, crypto, and child-process boundaries, (2) a third-pass IFC summarize / capture-origins / flow-shape wave, (3) shadow-daemon no-mock adoption + evidence-receipt execution layer, (4) bridge-2026 program decomposition, and (5) a wide reality-check defect cluster against the rewritten CLI/replay surface.

Delivered capability (sampled, by theme)

  • Theme 1 — TLS, crypto, child-process, Zlib. Peer-certificate metadata extractor with bounded X.509 DER parsing and Node-oracle fixtures (bd-il0d9: 5ee847e10, 46c8d93c3, cf3612917); RFC 6125 wildcard matching + root bundle + connect positional overloads (c7024687c); X.509 / IFC fix for TBSCertificate tag parsing + flow labels (08aa44884); Node-compatible process.nextTick queue without exposing raw process authority, required by stream fixture 0013 (bd-8nrud); P-256 ECDSA detached signing + DER verification
    • crypto.sign/verify lowering with IFC labels (42bd30256, a25c259e4, f7e303e6b); ed25519 sign/verify + generateKeyPairSync via typed RandomRead entropy + asymmetric key-lifecycle contract + zeroization lifecycle verification (fff2c4f7e, 7faab498c, 5fd27394b, 2c28f7113f); AES-256-GCM variable-IV transform with GHASH on decryption (dd251eba7, 99e5edf06); ChildProcess lifecycle bridge with bounded stdin + process effect errors retaining partial stdout/stderr on timeout / I/O / provider limits (bd-m42c2, bd-k709s); brotli runtime dispatch + sync-family closed-input flow + zlib / Buffer.equals IFC summaries (edd93e42b, 0cfb17480, ac689bc72, f4afa45d9, e9f929a4b, e2490c45d, 856a61de5); pull-based Brotli reader to throw on corrupt decompress input (89b926d4f, a6c2aea8e, b5fd26175 follow-up); brotli sync result → authenticated BufferObject classification (f4afa45d9).
  • Theme 2 — IFC summarize, capture-origins, flow shape. bd-dign3 partials 2-5 closed-result summaries for toString / valueOf / ArrayBuffer allocations / crypto metadata (58203b48f, e1362cf6d, 39696dc06, 4d0f63435); the new EngineParsedAggregate flow shape for builtin:QuerystringParse results, propagated through IFC summarization (80c26bf0f, 9d4874057, 151502548); ConstantsObject static-method + OwnKeyArray + matches! length/size recognition in flow labels (d8a84c418, c5b706838); FRANKEN_CAPTURE_DEBUG gate for capture-origins resolution (a7116c2fd).
  • Theme 3 — shadow + evidence layer. Authoritative shadow-daemon no-mock lifecycle drill (posture advisory-only, all five adoption gates green) — 698b9ae0a, 8b20e5f43, e8f58696e, 129a2f0f6, b5fd26175; execution-derived receipt layer with recomputable seals + receipt-integrity gate in main body + bd-q0cwt / bd-8hh0g close (ad43cd2a0, 675b3c232, 68690399b, 5ec3d7cb7, c432e530a); native promises in reaction completions, resolve, async returns (0bb43d4e7).
  • Theme 4 — BRIDGE-2026 program work. README structural-inventory drift + three nonexistent module references + repo hygiene (tu32j.1.8); restore repo-wide unsafe-code prohibition in test builds (tu32j.1.9); generate the live coverage matrix from authoritative beads / claims / harness inventory (tu32j.22.1.1); BRIDGE-21.61 enforces decomposed-parent closeout and exact required-child completion (4e37b213d).
  • Theme 5 — reality-check defect cluster (bd-sxh8o).* Promise/async then-handlers fix on frankenctl run + Array.prototype.toString prints [object Array] + benign frankenctl run reports containment_action=challenge
    • examples/21 verify.sh can green on usage errors + replay self-compares traces (bd-sxh8o.1..6, bd-25il9 was false-closed).

Plus the other open wave: ESM async module-graph evaluation with owned compiled IR per module, AsyncEvaluating dependency status, parent-import continuation, and rejection propagation for imported top-level await (bd-yn3lv); full set of [REALITY-2026-08-20] CLI defects resolved into the rewritten CLI; v0.1.0 release remains the only published GitHub Release.

Closed workstreams (sampled)

bd-53l89 (asymmetric key-lifecycle, phases 1-2), bd-acwe9 (zeroization lifecycle), bd-dign3 (crypto/IFC summarize, 5 partials), bd-nil04 (shadow no-mock drill, 5 gates), bd-q0cwt + bd-8hh0g (execution-derived receipts), bd-ifc-internal-label-cell-transcript-hplvg, bd-hwjbi (callback recursion), bd-fw7zd.9 / bd-fw7zd.13 (post-close / promise ownership), bd-z1peg (static IFC source labels), bd-z1peg.1..3 (authority/provenance registry + delegation accounting), bd-zco6t (fs compat residual), bd-8nrud (nextTick), bd-yn3lv (ESM async module-graph), bd-pafik (stream/promise provenance), bd-qjh7y (static IFC function summaries), bd-sblaq (lexical cells), bd-asw4m.3/.5 (lexical this / implicit net), bd-36e01 (Reflect.construct), bd-ur3tk.21 (PromiseStore), bd-i7cw8 (Readable reset), bd-sxh8o* (reality-check CLI cluster), bd-m42c2 (ChildProcess), bd-k709s (process effect errors), bd-lmchj (console hostcalls), bd-zyndq (BuiltinFunction new), bd-6o8kt (reduce BuiltinFunction), bd-25il9 (false-close correction), bd-mpm0l (per-iteration lexical), bd-es2ra (core for-of), bd-bscab (IR2 IFC multi-arg), bd-cbsgh (static reads on unforgeable brands), bd-22e1y (tls API fidelity), bd-2l1u2 (stream timers), bd-ur3tk (IFC register-label), bd-bscab (IR2 HostCall label join), bd-3gs6q (FullCapsHandler mock), bd-6a61n.2.7 (timer mock), bd-6wc97 (hostcall placeholders), bd-l4t7n (TLS), bd-h7p1a (capture-origins), bd-znj5l (Buffer.from over fresh aggregates).

Representative commits

  • 80c26bf0f — feat(ifc): add FlowValueShape::EngineParsedAggregate for querystring parse results
  • 151502548 — chore(engine,ifc): gated FRANKEN_CAPTURE_DEBUG diagnostic for capture-origins resolution
  • 08aa44884 — fix(tls): correct X.509 TBSCertificate tag parsing and IFC flow labels for observation methods
  • 42bd30256 — feat(crypto): implement P-256 ECDSA detached signing and verification with DER encoding
  • fff2c4f7e — feat(crypto): implement ed25519_sign_detached and ed25519_verify_detached with unit tests
  • 7faab498c — feat(crypto): ed25519 generateKeyPairSync via typed RandomRead entropy
  • dd251eba7 — feat(crypto): implement NIST SP 800-38D AES-256-GCM variable-length IV transform
  • 698b9ae0a — feat(shadow): authoritative no-mock lifecycle drill passes; no_mock_drill gate GREEN
  • ad43cd2a0 — feat(evidence): execution-derived receipt layer with recomputable seals and gate enforcement
  • 4e37b213d — feat(tracker): enforce BRIDGE decomposed-parent closeout via native gates
  • 5ee847e10 — feat(tls): add bounded X.509 DER metadata extractor (bd-il0d9)

Post-Snapshot Update — Janitor docs-reorg (2026-08-19)

Root planning and audit leftovers left the repository root. HASHER_AUDIT_FIXED_LAYOUT_MIGRATION.md and REVIEW_SUMMARY.md now live under docs/planning/. audit_observed_claims.py and test_script_fix.sh moved into scripts/. Skill-loop scratch is gitignored. This is hygiene, not a behavior change.

Representative commits

  • 49abf340d — chore(janitor): untrack skill-loop scratch; move root planning docs into docs/planning/
  • b1f5bc91c — chore(janitor): relocate remaining root reports and planning docs

Post-Snapshot Update — Sibling Dependency Isolation and a Standalone-Claim Correction (2026-07-25)

bd-ndpm2 was filed as a P0 build blocker: frankenengine-engine did not compile in any feature configuration because a half-finished async migration in /dp/frankensqlite broke fsqlite-btree. Re-measuring at HEAD found the originally-named crate fixed upstream and the same migration breaking one layer further down — /dp/sqlmodel_rust/crates/sqlmodel-frankensqlite, 32 errors. Chasing a sibling's in-flight refactor crate-by-crate is not a fix, so the engine's own coupling was cut instead.

Nine sibling path dependencies are now feature-gated, across three features that are all ON by default (a normal build in a full checkout is unchanged):

  • sibling-persistence — sqlmodel, sqlmodel-core, sqlmodel-frankensqlite, and through the last of those /dp/frankensqlite. cargo tree -e normal -i fsqlite proved that binding is the sole edge pulling fsqlite, so one cut drops two repos.
  • sibling-service-api — fastapi-core (gates policy_controller::service_endpoint_template).
  • sibling-dataframes — the five fp-* crates (gates the Parquet evidence-export lane).

typed_persistence_models.rs looked atomic — six #[sqlmodel(table = …)] model structs consumed by seven other modules — but TypedStoreRecord and TypedStorageAdapterExt turned out to be pure serde traits with no ORM coupling. Only the Model derive and the sqlmodel(…) attributes needed cfg_attr, so all six consumer modules were left untouched. Where a runtime path would otherwise degrade, it fails closed: Parquet audit export returns a typed error naming the required feature rather than emitting another encoding under the Parquet name.

New gate lane: ./scripts/test_standalone_build.sh sibling-isolation (also inside ci) asserts every /dp path dep is optional = true and that cargo tree --no-default-features -e normal,dev names zero /dp paths. It keeps enforcing under STANDALONE_BUILD_GATE_SKIP_REMOTE=1, since it needs no rch worker and the skip flag would otherwise disable the guarantee exactly when the heavy lanes are already skipped.

Two defects surfaced while gating. The self dev-dependency lacked default-features = false, so feature unification re-enabled default for every test target and cargo test --no-default-features silently rebuilt all nine siblings — the standalone-test lane had been proving nothing. The same leak existed on the three workspace crates that depend on the engine; fixing all four took the workspace sibling count from 119 to 0. Separately, two integration tests still asserted the removed fake FRANKEN_PARQUET_V1 header, directly contradicting the lib test that asserts its absence; both now assert real PAR1 framing.

README correction. "Standalone Mode (no sibling repos required)" was false and cannot be made true from inside this repository. Three experiments established that cargo resolves an optional dependency's manifest whether or not a feature activates it — for path, git, and unmatched [patch.crates-io] sources alike — and that the only source kind genuinely skipped when disabled is a registry dependency. None of the nine siblings is published. The README now states that --no-default-features links no sibling crates but still requires the checkouts present, and registry publication is tracked on bd-gw4cg. Claim-matrix spans were re-anchored and run_claim_to_proof_matrix_gate.sh ci passes (28 claims, 0 failures).

Post-Snapshot Update — Authenticated Process-Spawn Foundation (2026-07-23)

bd-x85a7 adds the first host process-execution authority to the extension host, as a fail-closed request/provider seam rather than an ambient capability. ProcessSpawn becomes a typed RuntimeCapability; ProcSpawnEffect carries a structured ProcessSpawnRequest instead of an untyped string tuple; and the default provider is DenyAllProcessSpawn, so a runtime that does not explicitly install a provider cannot spawn at all.

  • crates/franken-extension-host/src/process_spawn.rs implements the typed request/response/error protocol and a bounded native provider that re-verifies the target executable by SHA-256 before every launch, clears ambient environment and admits only an allowlist, jails the working directory to a canonical path, refuses dangerous environment carriers (LD_PRELOAD, DYLD_*, BASH_ENV), applies per-child resource limits, and returns opaque scoped handles rather than PIDs. A detached reaper thread plus Drop-time containment prevents orphaned children. The module keeps #![forbid(unsafe_code)].
  • crates/franken-extension-host/src/host_effect_journal.rs adds a globally ordered record/replay journal spanning host I/O and process effects. Reservation and completion preserve crossing order, because concatenating per-family transcripts would forge the ordering of interleaved effects and break replay identity.
  • The orchestrator gains set_process_spawn(...), a host_effect_journal result field, and last_failed_host_effect_journal; the capability stack treats ProcessSpawn as extraordinary authority that is absent from the ordinary Full profile, so it cannot be acquired by default composition.
  • Follow-up audit beads are tracked as bd-x85a7.1–.4: temporal containment and per-effect expiry, journal completion-hole integrity, secret hygiene for env/stdin/stdout/stderr in Debug and diagnostics, and platform process-tree containment (cgroup or job object) beyond Unix process groups. This entry records the foundation only; those four remain open.

Post-Snapshot Update — Generator Activation Boundary (2026-07-23)

bd-093id makes synchronous generator invocation run its parameter initialization before publishing the iterator, while suspending ahead of the first body statement. Generator activations now retain their full isolated execution context across yields; .next(value) injects the value and label at the prior yield site, returns completion records with done: true, and validates the generator receiver through the exposed next builtin.

  • Core IrSchemaVersion::CURRENT advances from 0.9.0 to 0.10.0 because serialized IR1, IR2 (through Ir2Op::inner), and IR3 gain the semantics-bearing GeneratorBodyStart variant. Schema gates now reject skipped peer-owned and future versions at lowering, verification, and execution entry points. IR1-to-IR3 lowering preserves a supported legacy header, so markerless 0.9.x generator artifacts retain their historical first-.next() start timing; current generator bodies must contain exactly one boundary before yield/return. Older readers must reject 0.10.x before decoding.
  • The engine mirror remains independently versioned at 0.7.0; this core-only checkpoint does not change the cross-seam ownership posture.

Post-Snapshot Update — Captured Local Lexical IR Metadata (2026-07-22)

bd-uhf1m preserves source let/const identity when a function-local binding is captured by a nested function and the parent body is detached for deferred IR3 lowering. Captured declarations now create correctly typed runtime cells, initialize them with InitBinding, and leave later assignments on the checked StoreScoped path. This restores const-assignment and temporal dead-zone behavior without changing synthetic, unresolved, inherited, or uncaptured bindings.

  • Core IrSchemaVersion::CURRENT advances from 0.8.0 to 0.9.0 because DeclareFunction and CreateFunction gain semantics-bearing persisted IR1 metadata. The optional vector is boxed so absent metadata does not enlarge the recursively lowered Ir1Op enum or consume an allocation; it is omitted when empty, so historical function-operation JSON and canonical bytes remain unchanged.
  • Core 0.9.0 readers accept historical core 0.8.x artifacts, whose missing metadata defaults to the legacy empty posture. Consumers must validate the header before decoding; older readers must reject a 0.9.x artifact rather than ignore the new metadata and reinterpret captured lexicals as mutable.
  • Core minors 0.6.x and 0.7.x remain rejected because they identify the incompatible engine-owned unresolved-name and follow-on Continue wires. The compatibility engine remains independently versioned at 0.7.0.

Post-Snapshot Update — Exact Module-Source Metadata (2026-07-17)

bd-lfq44 closes the remaining UTF-8 projection in parsed ECMAScript module sources. Imports and named re-exports now retain exact JsString code units through parser AST metadata, the engine parser arena, IR1, and IR3 constant pools. A source containing \uD800 therefore remains distinct from one containing \uDC00; neither can silently alias through a replacement-character projection during module lookup.

  • The compatibility engine AST schema advances from v3 to v4 and its IR schema from 0.2.0 to 0.3.0. The native core AST advances from v4 to v5 and its IR schema from 0.3.0 to 0.4.0. Both Cargo packages remain on the unreleased 0.2.0 compatibility-staging line.
  • ImportDeclaration::source and Ir1Op::ImportModule::specifier use JsString. Ordinary well-formed values keep their historical plain-string serde and canonical bytes; exact values use the established $wtf16 unit representation.
  • Named exports separate their canonical binding head from an optional exact module source. Source-free and well-formed clauses retain the historical scalar NamedClause payload. Only a non-well-formed source uses the namespaced $module_source payload, and readers reject that tagged form for well-formed content so the wire has one canonical encoding.
  • Runtime import dispatch carries the exact value to the existing filesystem conversion seam. A non-well-formed value is rejected there before path or cache lookup; the independent string-based resolver/ESM graph APIs are not reinterpreted or widened by this parser-path checkpoint.
  • Current readers retain same-major historical AST/IR inputs. Existing well-formed parser hashes and phase0 semantic artifacts remain unchanged; only schema identifiers, compatibility vectors, and IR-header snapshots advance.

Post-Snapshot Update — Parser EOF Coordinate Schema (2026-07-17)

bd-4tt6s corrects the canonical SyntaxTree root span in both parser seams. The root end_column is now the one-based UTF-8 byte column immediately after the original source on its final physical line, rather than a hard-coded 1. This covers single-line sources, non-empty multiline tails, trailing LF/CRLF, and exact byte widths for non-ASCII source text.

  • The compatibility engine AST schema advances from v2 to v3; the native core AST schema advances independently from v3 to v4.
  • The AST contract, serde shape, SHA-256 algorithm, and hash prefix are unchanged, so historical tree JSON remains readable. Historical v1/v2 engine version vectors and pre-correction hashes remain pinned explicitly.
  • Source-backed Parse Event IR materialization recognizes only the exact historical root-column defect, authenticated by its old tree hash; any other span or hash mismatch continues to fail closed.
  • Parser-generated canonical hashes intentionally change when the old root column was wrong. The live phase0 fixture catalog and its content-addressed artifact bundle are regenerated under the new schema checkpoint.

Post-Snapshot Update — 0.2.0 Compatibility Staging (2026-07-16)

bd-n8eta.4.6 moves only frankenengine-core and frankenengine-engine to an unreleased 0.2.0 line. This is a source-compatibility boundary, not a tag or publication: unrelated workspace packages stay at 0.1.0, and GA/release actions remain separately gated.

  • Both public baseline-interpreter Value enums are now #[non_exhaustive]. Downstream matches must retain a fallback arm before the separately owned typed-Symbol runtime work appends a new variant.
  • The downstream audit found 57 cross-crate match expressions across 13 files; every match already had a wildcard, binding, or equivalent fallback. /data/projects/franken_node had no direct construction, import, or match on either baseline Value type.
  • Existing serde discriminants and payload bytes do not change. Focused historical-wire tests decode and re-encode all pre-migration variants byte-for-byte; the later Symbol children own any new wire values and execution-seed state.
  • IrSchemaVersion::CURRENT remains 0.1.0 because this migration adds no IR opcode or persisted IR shape. The separately tracked bd-f1ixz must version its IR schema when its CopyDataProperties variants land on the unreleased 0.2.0 package line.

No Value::Symbol, typed executable property-key carrier, symbol_state, tag, or release is introduced here.

Core exact quoted-string schema slice (bd-vltnh)

The subsequent core-first bd-vltnh slice advances frankenengine-core::ir_contract::IrSchemaVersion::CURRENT to 0.2.0 and the native core AST schema to franken-engine.parser-ast.schema.v3. Quoted source literals, IR1 string literals, and the core IR3 constant pool now carry exact JsString values, including lone UTF-16 surrogates. Historical well-formed strings keep their prior leaf serde/canonical shape; exact values use the tagged $wtf16 unit representation. At that checkpoint the duplicated frankenengine-engine parser/IR mirror still remained at its prior schema; the engine section below records the subsequent mirror landing. Neither checkpoint is a release.

Engine exact quoted-string schema slice (bd-vltnh)

The engine mirror now advances IrSchemaVersion::CURRENT to 0.2.0 and its live canonical AST schema to franken-engine.parser-ast.schema.v2. Quoted expression literals cook directly into exact JsString code units, and that carrier is preserved through the parser arena, AST serde/canonical values, IR1 literals, IR3 constant-pool deduplication, and baseline LoadStr execution. Ordinary well-formed strings retain their historical JSON and canonical leaf shape; lone units use the tagged $wtf16 representation and remain distinct through an end-to-end source parse, lower, and execute cycle.

The later bd-lfq44 checkpoint above completes exact module-specifier metadata through the compiled parser/lowering path. Contextual legacy decimal escapes remain tracked by bd-xcqzp. Neither schema landing creates a tag or release.

Core CopyDataProperties IR schema slice (bd-f1ixz)

The core object-rest lane advances frankenengine-core::ir_contract::IrSchemaVersion::CURRENT to 0.3.0 and adds Ir1Op::CopyDataProperties { excluded_count } plus Ir3Instruction::CopyDataProperties { target, source, excluded, value_dst }. The frankenengine-core Cargo package remains on the deliberately unreleased 0.2.0 compatibility-staging line; this is an IR wire revision, not a package release or tag.

Existing IR variant names and payloads retain their prior serde and canonical representations. Core 0.3.0 readers accept historical 0.2.0 headers, while older readers reject the new externally tagged variants instead of silently reinterpreting them. The downstream source audit found only the engine differential oracle directly matching the core IR3 enum outside the core crate, and both matches already have fallback arms; /data/projects/franken_node has no direct imports, constructions, or matches of either core IR enum.

Post-Snapshot Update — Claim-Evidence Integrity Capstone (2026-06-21)

The CEI epic (bd-sde5e, stated state must be provably ≤ commit) reached its capstone. Tracks A, C, E, F were already closed; this update lands Tracks B and G and the Track-H reflexive claim:

  • B.2 (bd-sde5e.2.2) — every OBSERVED claim now carries a real verification_result=passed receipt from a live gate run (no backfill/pending). The last holdout, FE-CLAIM-022 (cross-runtime lockstep oracle), is re-emitted from a live run against real Node.js (/usr/bin/nodejs v20.19.4) with zero divergences across node/bun/franken; the latent bench bug behind its old backfilled receipt (the FrankenEngine lane exhausting the containment instruction budget on the 10k/20k-iter workloads) is fixed by sizing the comparative workloads to frankenctl run's real default budget. FE-CLAIM-023 (cross-platform identical-hash reproducibility) is honestly downgraded observed → target — it is unprovable on a single host and its receipt was backfill/pending (precedent: FE-CLAIM-012). Result: the A.3 bidirectional audit is 27/27 sound (100%), --blocking exit 0.
  • B.4 (bd-sde5e.2.4) — fresh-clone committed-evidence verification (scripts/e2e/fresh_clone_evidence_verification_smoke.sh) is 16/16 offline against a clean git worktree checkout.
  • G.1 (bd-sde5e.7.1) — scripts/run_claim_evidence_integrity_capstone.sh composes the four CEI checks (claim-to-proof wording + A.1/A.3 bidirectional lattice, H.1 Merkle ledger, whole-document consistency, D.3 Test262 posture) into one fail-closed meta-gate, with an e2e replay wrapper and docs/CLAIM_EVIDENCE_INTEGRITY_CAPSTONE_RUNBOOK.md.
  • G.3 (bd-sde5e.7.3) — a no-mock acceptance drill (scripts/e2e/claim_evidence_integrity_capstone_drift.sh) injects an over-promotion of each class, asserts the capstone reddens the responsible sub-gate, and restores the tree byte-for-byte: the gate cannot be satisfied by fixtures.
  • H.2 (bd-sde5e.8.2) — FE-CLAIM-025 reflexive soundness claim (the gate gates itself), backed by the A.5 adversarial corpus (4/4) and committed in the H.1 Merkle ledger alongside the claims it checks. The pre-existing unbacked target FE-CLAIM-TEST262 was also given a git-tracked evidence bundle so the blocking audit is clean.

Honestly deferred (genuine environment blocks, not skipped work): D.2 (bd-sde5e.4.2, full tc39/test262 corpus + denominator) and H.4 (bd-sde5e.8.4, a checked Lean 4 theorem) remain open; D.2 needs the external corpus and H.4 needs a Lean toolchain absent on this host. Their parents (Tracks D, H) and the epic stay open accordingly. Relates to the honest GA-exit ledger (bd-cixqu.47).

Commits: 619479f2, 163c17472.

Post-Snapshot Update — Track H Exception Semantics (2026-06-03)

  • Promoted the README JavaScript surface coverage for try / catch / finally and throw / exception semantics to Executed after the Track H closure-capture work (bd-cixqu.8.2, bd-cixqu.8.3, bd-cixqu.8.5, bd-cixqu.8.6, bd-cixqu.8.4).
  • Recorded try/catch/finally as a resolved parser and lowering inventory site, retaining FE-PARSER-GAP-TRY-CATCH-0001 only as historical fail-closed provenance.
  • Verified the exception semantics suite through rch: cargo test -p frankenengine-engine --test exception_semantics_conformance now covers 24 passing tests, including catch-binding closure capture and finally isolation.

Version Timeline

Kind distinguishes a published release from a plain git tag.

Version Kind Date Summary
main @ b1f5bc91c Unreleased main tip (not a tag, not a Release) 2026-08-19 Current window: crates.io sibling isolation, cell authority, RandomRead entropy, janitor docs-reorg.
0.2.0 (unreleased) frankenengine-core / frankenengine-engine Cargo versions 2026-07-16 → present Compatibility-staging line for public exhaustive-enum evolution; no tag or publication.
v0.1.0 Published GitHub Release 2026-05-29 First conventional release and installed-binary baseline. The only GitHub Release in this repo.
backup/main-tip-1b2e6cf0 Backup tag (not a release) 2026-04-16 Mid-April main tip preserved during the Test262 / async-execution work.
backup/worktree-tip-1f288b45 Backup tag (not a release) 2026-03-18 Mid-March worktree tip preserved during the integration-test enrichment wave.

The four chronological capability waves below are research-grouped, not release-tagged.

Per-Wave Metric Snapshot

Counts at each wave's closing commit, derived from git ls-tree -r --name-only <wave-end-sha> against the path patterns shown. Useful for "how much was built in this window?" without reading every commit. These are historical wave-end snapshots, not current-HEAD inventory counts.

Surface End of Wave 1 (2026-02-28) End of Wave 2 (2026-03-31) End of Wave 3 (2026-04-30) End of Wave 4 (2026-05-15)
crates/franken-engine/src/**/*.rs (recursive) 262 495 550 573
crates/franken-engine/tests/**/*.rs (recursive) 437 1,194 1,309 1,390
crates/franken-engine/tests/rgc_*.rs (top level) 0 34 36 37
scripts/run_*.sh (top level only) 118 201 227 241
.beads/issues.jsonl entries 951 1,118 1,739 2,584
Commits added in the wave ~315 ~670 ~2,115 ~1,346

Read across rows for "growth per wave": notice Wave 2's near-doubling of the test surface (437 → 1,194) under the iterator-protocol + exception-epic landings, Wave 3's surge in beads (621 new entries, from 1,118 to 1,739) under the claim-to-proof matrix introduction, and Wave 4's continued bead growth (+845 to 2,584) tracking the IDEA-WIZARD series.

Two of the README's "Code Surface At A Glance" counts use top-level-only patterns (src/*.rs and tests/*.rs, not recursive), which gives slightly lower numbers than the recursive wave snapshots; the recursive figures above include src/bin/, src/capability/, tests/_support/, tests/support/, and tests/conformance/.

Wave 1 — Bootstrap and scaffolding (2026-02-18 → 2026-02-28, ~315 commits)

The first ten days laid the entire repository skeleton: workspace structure, the canonical-encoding/IR/lowering/parser scaffolding, the original RGC ("Runtime Governance Compliance") gate framework, and the first wave of integration tests. The codebase grew from an empty repository to 262 source modules and 437 integration test files in this window (see the per-wave metric snapshot above).

Delivered capability

  • Cargo workspace with franken-engine, franken-extension-host, franken-engine-test-support, and franken-metamorphic crates plus an excluded in-progress franken-core extraction crate.
  • Repository constitution: AGENTS.md, docs/RUNTIME_CHARTER.md, docs/DONOR_EXTRACTION_SCOPE.md, docs/SEMANTIC_DONOR_SPEC.md — the binding rules that pin "native-only Rust core execution", #![forbid(unsafe_code)], the one-way franken_node → franken_engine dependency direction, and the claim-language policy that gates the whole project.
  • Parser + AST + multi-stage lowering pipeline (IR0 raw → IR1 normalized → IR2 simplified → IR3 executable), the original baseline interpreter, the execution orchestrator, and the evidence ledger — i.e. the architecture documented in docs/ARCHITECTURE_OVERVIEW.md.
  • First RGC gate framework: CI quality gates, artifact validator, runtime hotspot campaign, gate replay scripts, and the cross-track handoff protocol.
  • 40+ "expand core modules" passes covering canonical encoding, compiler policy, conformance catalog, capability witness, capability framework, security epoch, replay scaffolding, signed manifest, evidence emission, and the first revocation chain primitives.
  • First adversarial-testing surfaces: adversarial_coevolution, counterfactual_replay, tail_risk, bifurcation, rollback_synthesis, and the metamorphic-testing runner.
  • FRX (FrankenReact eXtension) track charters and the first FRX lockstep oracle / counterfactual evaluator.

Representative commits

  • 59a21498 — feat(engine): add RGC CI quality gates framework with tests, docs, and replay scripts
  • e308a853 — feat(engine): expand 40+ core modules with full runtime security, proof, and governance implementations
  • bd264466 — feat(engine): expand 8 core modules with canonical encoding, compiler policy, conformance catalog, etc.
  • f619b13b — test(engine): comprehensive integration test suite — 110 new test files
  • 639d8928 — feat(engine): milestone evidence gates, demo-claim linkage, flake quarantine, and swarm control loop
  • 4e344549 — feat(engine): adversarial coevolution, counterfactual replay, tail-risk, bifurcation, and rollback synthesis modules
  • fe44b2a3 — feat(engine): metamorphic testing suite — seed transcript logging + runner enhancements
  • dd2162d8 — feat(engine): add static semantics, TS module resolution, and RGC coordination modules
  • 02c47a89 — feat(engine): add RGC-063 cross-platform matrix verification contract
  • d99aec56 — feat(engine): enrich EvalError with correlation IDs, source locations, and stack frames

Wave 2 — Runtime semantics maturation (2026-03-01 → 2026-03-31, ~670 commits)

March was dominated by closing real JavaScript semantic gaps in the IR/runtime: the iterator protocol, exception/try/catch/finally lowering, generator/promise/spread semantics, ESM/CJS export resolution, module compatibility matrices, and a sweeping integration-test enrichment pass (the in-tree memory/enrichment_sessions.md records over 7,000 new tests landed in this window). It is also when the parser front-end picked up most of its ES2020 grammar surface and when the deterministic-hashing/length-prefixing audit hit most subsystems.

Delivered capability

  • Iterator protocol end-to-end: iterator_protocol.rs core substrate, 5 new IR1 opcodes (ForInInit, ForInNext, ForOfInit, ForOfNext, IteratorClose), real for..in / for..of lowering replacing the previous UnsupportedSyntax placeholders, and a 43-test iterable_workload_verification integration suite.
  • Exception epic (RGC-313): IR lowering for throw / try / catch / finally (extended BeginTry, added EnterFinally / EndFinally), a real runtime unwinder (CatchFrame, FinallyMode, pending_exception, real dispatch), rejection_reason_description in module rejection, and 13 exception-semantics conformance tests.
  • Generator + promise + spread semantics (RC beads 1.4, 1.12, 2.1).
  • IR3 instruction expansion: 19 new variants (Mod, Exp, Lt, Lte, Gt, Gte, Eq, StrictEq, NotEq, StrictNotEq, BitAnd, BitOr, BitXor, Shl, Shr, Ushr, InstanceOf, InOp, Construct) with matching baseline-interpreter dispatch and execution-orchestrator mnemonics.
  • ESM/CJS interop: overhauled ES2020 star re-export semantics, conditional and external exports map resolution, scoped-package and extensionless-relative tests, Node-compat CJS→ESM specimens, and a hybrid lane router; the module compatibility matrix gate gained npm-style pkg.js / @scope/pkg.js extension-probe anchoring with fail-closed package.json type=module behavior in native/node_compat modes.
  • Parser frontier (March-landed work only): initial tagged-template support and template-interpolation hardening, fail-closed handling for super / new.target / import.meta, named-export-clause validation, and the parser-oracle / parser-frontier-harness gates. (Earlier parser arena VariableDeclaration and named/namespace imports landed in Wave 1; tagged-template-as-Call, trailing-line-comment stripping, named-declaration export desugaring, simd_lexer content-binding, and the parallel scoped-worker lex all land later — see Waves 3 and 4.)
  • Cross-repo integration suite (scripts/run_cross_repo_integration_suite.sh) and machine-readable contract (docs/cross_repo_integration_suite_v1.json) for /dp/asupersync, /dp/frankentui, /dp/frankensqlite boundaries.
  • Deterministic-hashing audit: length-prefixing applied across content hashes (gate results, evidence bundles, signed manifests, rewrite packs, IFC declassification authorization, supremacy verdict aggregation, etc.) so concatenation collisions are no longer possible.
  • frankenctl CLI gained help <command> navigation, rch-wrapped replay command emission in artifact bundles, observability_mode JSON output and hash-stability regression tests, and preserved-bundle replay support across multiple gates.
  • Cache-oblivious metadata substrate, kernelized shift guard, semantic dark-matter engine (RGC-617), and the rough-path regime geometry orchestrator landed as performance/observability infrastructure.

Representative commits

  • 8753c439 — feat(engine): module system — async evaluation dependency tracking and compatibility matrix validation
  • f10150a2 — feat(engine): overhaul ESM export resolution to match ES2020 star re-export semantics
  • b17332ea — feat(engine): conditional exports map resolution tests for module resolver and compatibility matrix
  • 74a976ab — test(resolver): add scoped-package and extensionless-relative integration tests across compatibility modes
  • 44e1e65b — feat(engine): add cache-oblivious metadata substrate, kernelized shift guard, and semantic dark-matter engine
  • d7e1af52 — feat(engine): add rough-path regime geometry orchestrator (RGC-617)
  • bb8cb07f — feat(engine): cross-repo integration suite for multi-project contract verification
  • 135a1c2c — feat(rgc): add CI gate verdict, failure routing matrix, lane repro index, and health summary artifacts
  • 1bf264a8 — feat(engine): versioned rewrite pack — canonical pair keys, cost model guards, and pack diff
  • e8ed383c — fix(engine): deterministic content hashing — length-prefix fields and sort collections before computing digests
  • 215daf38 — fix(engine): seqlock fastpath — panic safety via SequencePublishGuard and poison recovery

Wave 3 — Real-execution conformance and benchmark truth (2026-04-01 → 2026-04-30, ~2,115 commits)

April was the most intense month by commit count and pivoted the project from "self-consistent infrastructure" to "real-execution truth". Test262 stopped using fake fixtures and started running real JavaScript; the benchmark harness stopped using hardcoded baselines and started measuring child wall-time and peak RSS via Linux pidfd+wait4; the claim-to-proof matrix v1 was introduced as the binding gate over every README claim; and the first set of "live" guardplane/IFC/quarantine examples replaced their mock counterparts.

Delivered capability

  • Claim-to-proof matrix v1 (docs/claim_to_proof_matrix_v1.json, docs/CLAIM_TO_PROOF_MATRIX_V1.md) wired into both scripts/run_claim_to_proof_matrix_gate.sh and the README. All 21 tracked claims classified as observed / target / hypothesis; the gate refuses progression when actual wording is stronger than the allowed state. bd-csnqb swept the unsupported "formal mathematical" claims and downgraded them in-tree.
  • Real Test262 harness: replaced fake test data and hardcoded fake results with actual JavaScript execution in the release gate; arrow-function output-mismatch regression closed; frontmatter parser hardened against overlapping markers; iterator-conformance comparison now uses the shared eval-vs-expected helper.
  • Real benchmark measurement: benchmark-e2e switched from timeout(1) shell wrapping to in-process timeout + threaded stdio capture, then to memfd-based stderr capture with OnceLock host-facts cache and typed artifact serializers; child wall-time and peak RSS measured via Linux pidfd+wait4 with stderr timing-footer portability fallback; live Node/Bun baseline measurement (bd-16ch6); hardcoded throughput baselines eliminated (bd-1pq04); fake containment-latency data eliminated (bd-69kbi); cross-runtime output equivalence proved from captured bytes.
  • Real interpreter semantics: Array.from(iterable, mapFn, thisArg); Generator/Async/AsyncGenerator dispatch; Function.{call,apply}; reduceRight; Map/Set/WeakMap/WeakSet seeded from iterables; Promise.all delegated to combinator; async function execution semantics (bd-2lg6f); receiver-aware builtin dispatch with real Array.some callback; full function-body try/catch/finally + JumpIfFalsy two-target lowering + EnterCatch label binding; IR3-aware eval completion; IR3 TemplateLiteral emission; SharedBudgetEnforcer so subsystems observe live certificate updates; GovernanceContext composition root (bd-2hzkh).
  • Parser front-end finalization: tagged-template expressions parsed as Call with template-literal argument; trailing line comments stripped and unseparated expression sequences rejected; named-declaration exports desugared into declaration + named clause; same-line statements after export function / export class blocks split correctly; simd_lexer token witnesses mix input hashes so token outputs are content-bound.
  • Live "impossible-by-default" examples: live guardplane posterior + expected-loss decision example; live quarantine propagation with convergence evidence; live IFC/declassification example with signed receipts (bd-dpfvh); live capability rejection example (bd-1bao8); production_feature_catalog gate companion; bench-vs-node example (bd-79rwx); certified rewrite optimization demo; react compile demo (bd-3eydu); decision-receipt demo. All 13 "impossible-by-default" capabilities now have demo directories (one without a dedicated directory; see examples/README.md).
  • Proof-artifact contract (proof-artifact module): shared manifest module + script helper; adopted across three existing gates; events.jsonl race condition fixed with atomic emission; enumeration validation with Ord trait; cryptographic content binding added to IFC system.
  • Red-team / attacker harness (bd-28otw): attacker execution harness with explicit scenario outcomes replacing hardcoded baseline assumptions; comprehensive baseline validation in compromise-rate gate.
  • Fuzzing: parser fuzz harness, proof-artifact JSON validation targets (PHASE 3), shadow_panel_bundle target (bd-hbil1), ts_module_resolution_resolve target (bd-6fcpn), parallel-parser coverage-guided fuzz harness.
  • Shadow daemon adoption gates + mutation policy enforcement — preserved the advisory-only mode invariant documented in docs/SHADOW_DAEMON_PROOF_STATE.md.
  • Privacy verification artifacts, declassification timestamp bounds, replay coverage proof metric gate (bd-2488a), throughput disruptive-floor metric gate with Node/Bun denominators, three new metric gates and one proof example (bd-38mby, bd-1qr4f, bd-3mp80).
  • frankenctl run subcommand expanded with structured output + capability flags; --observability-mode consistently surfaced in JSON; the README CLI smoke workflow (scripts/e2e/readme_cli_workflow_smoke.sh) was wired to the shared proof contract (bd-1fjqa).
  • GA exit evidence package, cross-architecture reproducibility contract, deterministic support-bundle export, workload preflight doctor workflow, deterministic technical-report renderer, rollout-controller guardrails, replication claim tracker, acceptance ledger.

Representative commits

  • afe84382 — feat(claim-matrix): seed claim-to-proof matrix v1 + wire into gates and README
  • 71cda5e5 — feat(claim-matrix): add gate runner script + tighten must_contain anchors
  • e84796f1 — feat(bd-csnqb): audit and downgrade unsupported formal mathematical claims
  • d21262af — feat(test262): implement real Test262 conformance harness integration
  • 21b485a0 — feat(test262): replace fake test data with real JavaScript execution in release gate
  • d728d81a — feat(benchmark-e2e): measure child wall-time and peak RSS via Linux pidfd+wait4 with in-band stderr timing footer fallback
  • 38cfc002 — feat(benchmark-e2e): memfd-based stderr capture, OnceLock host-facts cache, runtime launch resolution, and typed artifact serializers
  • f5847faa — feat(bd-16ch6): implement live Node/Bun baseline measurement for throughput gate
  • 30f3aa96 — feat(bd-1pq04): eliminate hardcoded throughput baselines and add defensive validation
  • 4b2c2b03 — feat(bd-69kbi): eliminate fake containment latency data and add defensive validation
  • d2ea4d17 — feat(examples): implement bd-dpfvh live IFC/declassification example
  • ab686dfa — feat(proof): Live quarantine propagation example with convergence evidence
  • 267095c8 — feat(examples): implement bd-1bao8 live capability rejection example
  • 029e9454 — feat(proof): Live guardplane posterior and expected-loss decision example
  • 242bf0b5 — feat(replay): add replay coverage proof metric gate (bd-2488a)
  • 8086b135 — feat(metrics): Implement throughput disruptive-floor metric gate with Node/Bun denominators
  • 9e3576b1 — feat(metrics): implement bd-1vwza compromise rate metric gate
  • aa11e88c — feat(baseline-interpreter): add Generator/Async/AsyncGenerator dispatch + Function.{call,apply} + reduceRight
  • 5a3d047a — feat(lowering): function-body try/catch/finally + JumpIfFalsy two-target lowering + EnterCatch label binding
  • d728d81a — feat(benchmark-e2e): wall-time and peak RSS via pidfd+wait4
  • 39ded447 — feat(red-team): implement attacker execution harness (bd-28otw)
  • f76c92e9 — feat(fuzz): add parallel parser coverage-guided fuzz harness
  • 8c5c9459 — feat(proof-artifact): Fix events.jsonl race condition with atomic emission
  • 3cb9c7a8 — feat(governance): add GovernanceContext composition root (bd-2hzkh)

Wave 4 — Claim promotion, proof-specialized optimization, rch hardening (2026-05-01 → 2026-05-15, ~1,346 commits)

May has been about turning observed-but-fragile gates into hard contracts. The "idea-wizard" series (X, XI, XII, XIII) walked the README's remaining hypothesis claims toward observed by adding explicit proof bundles, rollback receipts, and no-mock acceptance drills. Concurrently, the rch (remote compilation hooks) infrastructure was hardened against worker drift, shard pressure, and brownouts so the large-batch agent swarms could keep landing work. The franken-core extraction crate started executing real class semantics for the first time.

Delivered capability

  • franken-core extracted runtime modules finally land in a compileable form (bd-zsais); class semantics start executing for real — class-expression semantics, extends/super dispatch, new.target in constructors, accessor getter/setter descriptor invocation via GetProperty/SetProperty, baseline-interpreter execution of class accessor get/set descriptors, private-accessor-key prefix tagging during lowering, heap-backed own-property storage for callable values + class-lowering Pop fix.
  • Async generators: .next() body execution implemented; async function execution semantics finished (bd-mw20e.2); pending-await contract made explicit (bd-jcqqj follow-up); await IFC labels preserved (bd-jcqqj).
  • Proof-specialized optimization promotion control loop (IDEA-WIZARD-XI, parent bd-xg3d6): promotion-control contract & inventory (bd-sisok), deterministic promotion eligibility composer (bd-4j2ck), demotion rollback and safe-mode replay receipts (bd-or2e1), workload-regime transfer guard for promotion decisions (bd-jp4r0), promotion-state surfacing in operator runbook/status (bd-yo0eh), no-mock promotion-control replay drill and truth gate (bd-xbesa).
  • Real hot-path proof (IDEA-WIZARD-X, bd-t5k40): rejected simulated hot-path evidence; rch hot-path wrapper (scripts/run_real_hot_path_proof.sh smoke); hot-path evidence runbook (docs/REAL_HOT_PATH_EVIDENCE_RUNBOOK.md); hot-path contract goldens; hot-path evidence drill. FE-CLAIM-010 (Node/Bun denominator) explicitly kept target until live denominator artifacts replace placeholders; MockCertificate and hot_paths_simulation artifacts now rejected as backing evidence.
  • Zero-ready validation truth lane (IDEA-WIZARD-XII, bd-n51l8): rch policy gate is now wrapper-aware (bd-n51l8.1); closed-bead semantic contradiction scanner (bd-n51l8.2); reopen real pending-promise await execution from source evidence (bd-n51l8.3); zero-ready source-gap picker for bounded follow-up beads (bd-n51l8.4); zero-ready truth surfaced in operator handoff status (bd-n51l8.5); no-mock drill for the lane (bd-n51l8.6).
  • README hypothesis-claim promotion (IDEA-WIZARD-XIII, bd-ly6hp): claim-promotion contract for hypothesis gaps (.1); transparency-log decision receipt proof bundle (.2); live quarantine mesh convergence proof (.3); capability-typed ambient-authority rejection pilot (.4); README claim promotion gated on live proof artifacts (.5); no-mock claim-promotion acceptance drill (.6); explicit rollback-evidence requirement in the promotion gate runner (bd-zso7f).
  • rch (remote compilation hooks) brownout/preflight hardening: worker-pressure preflight is now default; rejection of expected/native/selected worker drift; route preference propagation; preserved worker status on route drift; fail-closed lib-unit smoke execution; rch policy-gate awareness of wrapper cargo gates; shard pressure preflight contract; fail-closed shard runner; shard-runner termination classification; opt-in shard keepalive instrumentation; brownout source closeout + validation baseline documentation.
  • Parser & lexer: parallel lex chunks execute on scoped workers; continued logical-line indentation normalization; debug-derive for scoped chunk lex; trailing line-comment stripping; same-line statement splitting after export blocks.
  • Re-enabled certified_rewrite_optimizer (3f046a2a): aligned with current APIs after a long pause.
  • Shell hygiene smoke gate (bd-j2o4x): matrix coverage of operator and e2e scripts.
  • Topology queue admission decisions, removal of placeholder authenticity-signature seeds, hardening of proof_release_gate (requires cas:// URIs to embed archive_root hex prefix), and continued length-prefixing pass across remaining hash inputs (conformance harness failure-id, repro-digest, React package cohort, hole witness generator, hardware parameter manifold, AOT entrygraph compiler, GovernanceReport schema/spec, support-bundle evidence, GateResult variable-length fields).

Closed workstreams (selected)

  • bd-xg3d6 (IDEA-WIZARD-XI parent) — Proof-specialized optimization promotion control loop
  • bd-ly6hp (IDEA-WIZARD-XIII parent) — Promote README hypothesis claims with live proof bundles (parent open at time of writing; children .1–.6 closed in sequence)
  • bd-n51l8 (IDEA-WIZARD-XII parent) — Zero-ready validation truth and semantic debt control plane
  • bd-t5k40 (IDEA-WIZARD-X) — Replace simulated hot-path evidence with real runtime proof lanes
  • bd-2488a — Replay coverage proof metric gate
  • bd-1vwza — Compromise rate metric gate
  • bd-38mby / bd-1qr4f / bd-3mp80 — Three named metric gates landed with proof examples
  • bd-zso7f — Explicit rollback evidence requirement in promotion gate runner

Representative commits

  • f925fcf5 — feat(proof): add optimization promotion contract
  • 511f02a1 — feat(proof): add optimization promotion composer
  • ecb397fe — feat(proof): add optimization demotion receipts
  • c695936d — feat(proof): add optimization transfer guard
  • df1f3121 — feat(proof): add optimization operator status
  • 42428d68 — feat(proof): add optimization replay drill
  • 636d7f11 — feat(proof): add rch hot path wrapper
  • 31798477 — test(claims): reject simulated hot path evidence
  • f4b0e27c — docs(proof): publish hot path evidence runbook
  • 177ddc52 — test: add quarantine mesh proof wrapper
  • c4c350f6 — test: add transparency receipt proof bundle
  • ab19aa69 — test: add capability typed authority proof
  • 8c6a4038 — test: add claim promotion contract gate
  • 6b004fdc — test: gate xiii claim promotion reports
  • 2b7886de — test: add xiii claim promotion acceptance drill
  • d37aa248 — fix: require promotion rollback evidence
  • d51f2715 — test: add shell hygiene smoke gate
  • 32574c81 — feat(baseline-interpreter): implement async function execution semantics (bd-mw20e.2)
  • 9611a028 — feat(async-generators): implement async generator .next() body execution
  • 3f046a2a — feat(franken-engine): re-enable certified_rewrite_optimizer module and align with current APIs
  • 9512282b — feat(franken-core): land the five extracted runtime modules to make standalone manifest compileable (bd-zsais)
  • d35c9758 — feat(franken-core): execute class extends super dispatch
  • 7af839bc — feat(franken-core): execute new.target in constructors
  • b5d4aae6 — feat(parser): execute parallel lex chunks on scoped workers
  • 7a16247f — feat(rch): add shard pressure preflight contract
  • bc84e2cf — feat(rch): add fail-closed shard runner
  • b696d496 — feat(audit): objective artifact completion audit gate (new contract)

Cross-cutting workstreams

These are visible across all four waves and worth tracking separately from any single month.

Beads as the unit of work

The project uses br (the Rust-port beads_rust tracker) with issues checked into .beads/issues.jsonl. The README's claim-language gate (docs/CLAIM_TO_PROOF_MATRIX_V1.md) names a specific owning bead for every tracked claim. The IDEA-WIZARD-* series above each correspond to a parent bead with .A–.F children; the same shape recurs across earlier RGC-* epics.

Always-on gates layered onto every change

scripts/ ships 241 run_*.sh files. Major families:

  • run_rgc_* — Runtime Governance Compliance (56 scripts: cross-platform matrix, security enforcement, runtime semantics, statistical validation, performance regression, JSON compound traversal, NPM compatibility matrix, observability publication policy, module interop matrix, CLI operator workflow, docs/help surface audit, zero-placeholder, etc.)
  • run_parser_* — Parser (36 scripts: oracle, phase0 artifact, performance promotion, frontier harness, operator runbook, gap inventory)
  • run_frx_* — FrankenReact/FRX (34 scripts: canonical React corpus, SSR/hydration/RSC, local semantic atlas, Track D WASM lane, Track E verification/fuzz, online regret + change-point demotion controller)
  • run_claim_to_proof_matrix_gate.sh, run_real_hot_path_proof.sh, run_reproducibility_contract_suite.sh, run_metamorphic_suite.sh and other top-level claim/evidence gates.

Every gate has a matching scripts/e2e/*_replay.sh wrapper that can replay the latest preserved artifact bundle or a pinned timestamp under artifacts/.

Determinism and length-prefixed hashing

A repeating motif across all four waves: any time a content hash mixes variable-length data, the change replaces concatenation with length-prefixing so that distinct field decompositions cannot collide. Examples appear in February (initial scaffolding), March (rewrite packs, support bundles), April (proof-artifact contract), and May (GateResult variable-length fields, AOT entrygraph hashing, conformance-harness failure-ids).

Sibling-repo reuse contract

Per docs/RUNTIME_CHARTER.md §5, franken_engine consumes (one-way): /dp/asupersync (control plane), /dp/frankentui (TUI), /dp/frankensqlite (persistence; with the documented DEVIATION for typed-heavy stores still routing through generic storage_adapter.rs), /dp/sqlmodel_rust (typed schema layers, partially adopted), /dp/fastapi_rust (service/API control surfaces). Cross-repo integration is verified by scripts/run_cross_repo_integration_suite.sh and pinned in docs/cross_repo_integration_suite_v1.json.

Notes for Agents

  • The version timeline is sparse on purpose: the only GitHub Release is v0.1.0 (2026-05-29). Cargo 0.2.0 is an unreleased compatibility-staging line, not a tag. To pick an "as-of" snapshot after that, use a commit SHA.
  • The capability-wave sections are research-grouped by month because that's how the work actually landed; they are not release notes.
  • For any claim made in README.md, the authoritative gate is scripts/run_claim_to_proof_matrix_gate.sh ci against docs/claim_to_proof_matrix_v1.json. If a README claim says something stronger than the matrix allows, the gate fails — the matrix wins.
  • For implementation evidence behind any wave entry, follow the linked commit, then git log -- <touched-files> from that commit to see the surrounding cluster.
  • For workstream intent, look up the bead id (bd-xxxxx or bd-xxxxx.N) with br show <id> or read .beads/issues.jsonl directly.
  • Per-session enrichment notes (large test landings) are kept in memory/enrichment_sessions.md, memory/completed_beads.md, and the dated session files (memory/session_*.md). Treat those as frozen-in-time field notes, not current state.
  • legacy_v8/ and legacy_quickjs/ are reference corpora only; the runtime charter forbids them from becoming runtime dependencies.