promotion UX: support partial / edited promote (extract knowledge from hybrid noise+signal rows)

[tznthou]

Context

v0.3.0 promotion path (POST /journal/promote/:id) copies the entire session_journal.content field into memories.content — promote is row-grained. After 5/7 dogfood, cross-provider review (Codex, 2026-05-08) flags row granularity as a hard limit on journal yield.

Finding

Journal entry id=21 (5/7 capture) demonstrates the hybrid row problem:

```
Content opening (visible in pending preview, first ~200 chars):
雙寫同步、diff 空、139 行(~150 軟上限內)。
[wrap-up summary]
更新檔案: ...
本次 archive: ...
Promote 候選 (待子超 confirm):

Content middle (chars ~800-1500):
ImageMagick inline chroma key + G despill 一條 magick 命令版
(fuzz 25% transparent + ( -clone 0 ... ) parens stack + CopyGreen composite)
zsh 對 +clone[0] 報「no matches」要 bash subshell +
ImageMagick 的 [N] 只能用在 input filename 不能用在 +clone[N]
```

This row is noise wrapper + signal core:

• Wrapper layer (lines 1-30): wrap-up summary, file diff, archive list, promote candidate list — all transient
• Core layer (lines 30-50): ImageMagick chroma key command, zsh syntax pitfall, +clone[N] limitation — durable technical knowledge

Current options under v0.3.0:

• ccmem promote 21 → memories pollutes with wrap-up text (defeats trust split)
• ccmem reject 21 → ImageMagick / zsh lessons silently lost

Why this matters

• Hybrid rows are not edge cases — any session that records a real outcome and then summarizes will produce one. 5/7 had 1/16 such row in 24h.
• The trust split (v0.3.0 P1) is structurally vulnerable: pending → promoted → memories preserves the pollution if promotion is row-grained
• Codex framing: "promotion UX must support extract / edit promote, otherwise journal can isolate but cannot transform"

Proposed fix

Mode A (minimum viable): ccmem promote <id> --content "..."

Accept stdin or --content flag with edited body; the new memory's content is the supplied text, but source_session_id and source_journal_id (new column) preserve traceability back to the original journal row.

Mode B (better UX): ccmem promote <id> --extract

Spawn $EDITOR with the journal row pre-loaded; user edits to keep only the durable signal; on save, the edited buffer becomes the memory content.

Both modes share:

• memories.source_journal_id foreign key to journal entry (new column, nullable for legacy memories)
• Original journal row retains status='promoted' and references the resulting memory id
• content_hash on memory recomputed from edited content (avoid hash collision with raw row)

Acceptance criteria

• POST /journal/promote/:id accepts optional body.content parameter; if present, memory.content = body.content (else current behavior preserved)
• ccmem promote <id> --content "edited text" CLI path
• ccmem promote <id> --extract opens $EDITOR, captures saved buffer
• memories.source_journal_id migration (schema v23)
• Test cases: hybrid row id=21-equivalent → extract → memory content is core only, journal row marked promoted with reference
• Backward compat: existing ccmem promote <id> (no flags) still copies whole row

Risk

• $EDITOR spawning in CLI requires TTY detection — not all environments have it (CI, automation). Fallback: --content mode or error message
• Trace integrity: if user edits content beyond recognition, source provenance becomes fuzzy. Acceptance: source_journal_id is provenance not validation
• Schema migration v22 → v23 needs same care as v22 (backup before, integrity check after)

Related

• Surfacing UX: SessionStart hook nudge for pending journal review #28 (open) — surfacing UX; pending list preview also affects which rows look extractable. Codex flags pending listing newest-first + 200-char preview as systematically biasing review against rows like id=21 (wrapper opens, signal trails)
• split low-trust session_journal from high-trust memories; add access-based decay #21 (closed by PR P1 (#21): split low-trust session_journal from high-trust memories #26) — v0.3.0 trust split that ships row-grained promote path
• Make tags first-class in recall — extend to memories table and add filter to recall_query #15 (open) — tag first-class; relates to whether extract should also support per-span tagging (deferred)

Cross-provider review trail

Codex review (2026-05-08): row granularity is a structural limit on journal yield; promotion UX without extract / edit is the bottleneck after v0.3.0 fixes the persistence gate.