[idapixl]

Why

#85. In an attended dream on 1.10.0, the signature gate reported clean while refine made two damaging rewrites:

1. A paper summary joined an unrelated memo-idea memory at cosine 0.712. Refine appended the summary verbatim and dropped the memory's opening "Virgil's memoir-agent idea:" (688 → 1,177 chars).
2. Another memory lost its closing "(v0.8.1)".

Only hindsight called checkRewrite. Its number check also missed version strings, because \b sees no boundary between v and 0.

What

• Refine runs checkRewrite with a new limitGrowth option: a definition that grows past max(1.4×, +300 chars) is declined. Declined rewrites keep the old definition, write no belief row, and are counted in phases.refine.declined.
• New guard checks for every caller:
  • dropped named entities (capitalised words that appear mid-sentence);
  • a lost opening subject;
  • entity comparison by whole word, ignoring case;
  • version strings in the number check.
• Cluster join floor goes 0.70 → 0.75, configurable with the namespace's cluster_threshold. dream and auto-consolidation now pass it through. A wrong join hands refine unrelated text; a missed join only mints a separate memory.
• Abstract names an unlabelled synthesis with deriveName, the way create names memories, instead of clipping its first clause.
• dreamConsolidate now reports phases.hindsight.declined. It was counted but left out of the result.

Verification

• 459/459 tests. Of those, 12 are new (6 guard units, 6 refine/cluster/abstract end-to-end), and 9 of the new ones fail on master. The fixtures are the two real rewrites from the run.
• I replayed the guard against a real store's full belief history (233 rows). The 20 rows declined only by the new checks are all inflated stubs, truncations, or generalisations that dropped their specifics.
• I measured on a copy of the same store with a packed build of this branch:
  • the paper became its own memory;
  • the v0.8.1-dropping rewrite was declined;
  • 0 damaged rows, 0 regressions, 0 belief rows for the diff to flag, salience ≤ 1;
  • abstraction names are concept labels.

Not in this PR

• Abstraction content quality: two of three abstractions in the measured run restate one theme through loose cross-domain glue.
• A declined refine consumes its clustered observation: cluster marks it processed before refine runs, so its content reaches no memory.

Both are filed separately.

Closes #85

🤖 Generated with Claude Code