chore(perf): retire scripts/perf/evidence into ledger bullets (#4117)

87 files and 2.3 MB of rejected-experiment blobs and unrun harnesses, none of
which CI executes and none of which any scripts/perf entrypoint reads. The tree
needed a CI bypass to exist: every one of the seven evidence-only PRs that built
it (#4060, #4061, #4068, #4069, #4070, #4071, #4072 -- each touching nothing but
the tree and the ledger) carried revert-oracle-exempt, which skips the lane
entirely. #4069's body says why: "The oracle classified retained measurement JSON
as production code and failed because it has no changed application test." Three
of its READMEs specify zero-context patches so blank context lines do not trip
the whitespace gate. The repo keeps what CI runs and what the next agent must
read before spiking; the rest lives in git history at a named SHA.

Each of the nine ledger sections that linked into the tree becomes one bullet in
the section matching its verdict, in the house style the older #1445 entry
already uses: verdict, the mechanism that failed, the headline number with its
cohort size, the issue and PR, and a `git show` pointer pinned to
4fbbe8d. Five went to "Dead ends", four to
"Shipped wins", and the #3978 harness smoke folded into the manual-readiness
section as one sentence. #4031's verdict joins them: a dead end that was
recorded only in an issue comment and absent from the ledger entirely.

Two things stay inline rather than behind a pointer, because they are what a
re-spike gets wrong: the PGO RUSTFLAGS decisions (empty control, -Cprofile-
generate with its Darwin section alignment for counter-only training only,
-Cprofile-use, and the build-std/target split that makes the probe's profile and
the server's non-interchangeable), and the Y-up orientation rule
(x, y, z) -> (x, z, -y).

The archived patches are the only public copy of their mechanism. d979e92e4,
3e675edea, 67c3f6d31 and bdc38d30c are on no remote (`git branch -r --contains`
is empty for all four), so each bullet cites its patch path plus a public apply
base checked reachable on origin/main: 96ea5f0, e409924, 1b95c66. The two
chained follow-up patches (later-tests.patch reproducing bdc38d30c,
test-followup.patch reproducing c0ef3e802) get a pointer each for the same
reason. All five patches were applied for real against their cited bases with
`git apply --cached --unidiff-zero` into a temporary index; all five succeed and
both follow-ups chain after their measured patch. 7509432 is NOT on
origin/main -- it is the local pre-squash measured source -- so it is not cited
as an apply base. Every one of the 15 `git show` pointers resolves.

.github/workflows/test.yml's path-filter comment names the tree. The claim is
historical and stays true, so it is only marked as removed so a future grep does
not chase the path. No workflow logic changes. `git grep perf/evidence` now
returns that comment and the pinned pointers, nothing else.

Deleting the tree also removes the only Python project root outside
rust/python and tools/ifcopenshell_reference, and with it the dependabot noise
source behind #4073 and #4074.

Closes #4112.


Claude-Session: https://claude.ai/code/session_0193douQ6sTYHE65DJmyAei9

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: louistrue

.github/workflows/test.yml

@@ -276,10 +276,11 @@ jobs:
             # scripts/lib/revert-oracle.mjs byte-for-byte (any `test_*.py` /
             # `*_test.py`, any depth) rather than naming known marker
             # directories by hand, so a NEW Python test tree can never fall
-            # outside this filter the way rust/python's and
-            # scripts/perf/evidence's Python test fixtures did before this
+            # outside this filter the way rust/python's and the since-removed
+            # scripts/perf/evidence tree's Python test fixtures did before this
             # fix — both already existed unfiltered when this comment was
-            # written.
+            # written. That evidence tree was retired in #4112; the path no
+            # longer exists and nothing needs to be looked for there.
             python:
               - '**/test_*.py'
               - '**/*_test.py'