typescript-decompiler/docs/phase-docs/phase-7-iterative-relabel-evaluation-application-and-queue-feedback.md

Phase 7 — iterative relabel evaluation, application, and queue feedback

Goal

Validate Phase 6 model responses, evaluate and apply accepted names safely after wave reconciliation, and feed accepted-name results back into queue state so Phases 5 through 7 can iterate together.

Script

scripts/apply-relabel-results.js

Inputs

• Phase 5 queue state from runs/<run-id>/queue.sqlite
• executed Phase 6 batch artifacts such as runs/<run-id>/batches/<batchId>/request.json and runs/<run-id>/batches/<batchId>/response.json
• fixed type-specific or pass-specific response schemas emitted by Phase 5
• configured confidence thresholds, including any per-pass-kind overrides
• current generated source files or source metadata
• reusable naming memory state

Safety rules

• only apply renames for bindings proven local to the segment scope
• keep original identifiers in metadata when useful for traceability
• reject reserved words, invalid identifiers, unchanged names, and local collisions
• validate every model response against the fixed schema for that candidate type or pass kind before using it
• enforce deterministic machine-checkable naming rules derived from the naming-conventions prompt constraints
• reject insufficiently specific names before collision resolution when specificity checks fail
• do not invent fallback names or auto-adjust model proposals

Iteration rules

• reconcile results at the wave boundary; batches in the same wave must not affect each other mid-wave
• accept names independently when the model is confident about one symbol but uncertain about related symbols in the same response
• support partial attempts inside one work item, because a batch attempt does not imply every candidate in that item was meaningfully attempted
• track candidate naming attempts separately from batch attempts and insufficient-evidence counters
• immediately apply accepted names only after wave-level reconciliation has produced the final accepted set for that wave
• write accepted-name feedback back into queue state so Phase 5 can recompute evidence, difficulty, and priority deterministically on the next iteration
• send unresolved candidates to the back of the queue instead of forcing a guess
• stop iterating when all candidates are named, all remaining candidates hit configured retry limits, or no new names are being accepted

Outputs

• updated runs/<run-id>/segments.jsonl
• updated generated source files or source metadata
• updated runs/<run-id>/queue.sqlite with candidate results, statuses, defer reasons, attempt counts, insufficient-evidence counts, and accepted-name feedback
• preserved executed batch request and raw response artifacts for replay and audit
• updated stable/naming-memory.json

Release-oriented requirements

• relabel output should improve browsing and editing, but not become a hidden semantic transform layer
• naming memory should remain small and reviewable

Verification

• validate each response file against the fixed schema expected for its batch or work-item type
• confirm candidate naming attempts, batch attempts, and insufficient-evidence counters are recorded separately
• confirm wave-level reconciliation happens before any names are applied
• parse generated output again with Babel
• confirm no syntax breakage
• confirm accepted names are written into naming memory for future reuse
• confirm accepted-name feedback is persisted to queue state for the next Phase 5 reranking step
• confirm low-confidence, insufficient-evidence, insufficient-specificity, and collision-risk candidates are requeued or terminally marked with correct reasons