feat: batch feedback-ingestion mode (the 44-items problem) under the amend scope

[SaRedfiche]

Problem

When a user finishes a run and finds a batch of missed items (a customer reported 44: missed requirements + code-review comments), there is no way to feed that batch back into the in-progress or completed intent. One-at-a-time Redo or a fresh flow is the only path, which is why a trivial 11-line fix cost ~2 hours and 21 prompts.

Proposal: batch feedback-ingestion mode (runs under the amend scope)

Let a user hand the workflow a batch of review comments / missed requirements, and have AI-DLC triage them, plan a minimal set of stage/file touches, and drive them to verified completion in one governed pass.

• Entry: /aidlc amend --from-review <path-or-inline> and /aidlc amend --from-requirements <path> (flags land with the amend scope). Accept unstructured (pasted PR comments / checklist / scan report) and structured (a findings file) input.
• Ingest & normalize into discrete findings {id, description, type: missed-requirement|review-comment|defect, targetHint?}.
• Triage & cluster by the artifact/stage each implicates (Requirements / Design / Code-Gen / Build&Test); dedupe; produce a per-stage approval table (mirroring the composer's decision table) with a reason per finding — approved before any change.
• Plan a minimal touch set — an EXECUTE/SKIP plan over the existing workflow's stages that enriches affected artifacts in place (Goal 6 / [Feature]: Propagate build-time discoveries (rate limits, API quirks, pivots) across units, independent of declared dependencies #299 direction), bounded (cap stages touched; surface the count up front).
• Drive to completion — execute approved touches, run the Verifier on changed artifacts only, route test failures through the feat: bounded Build & Test → Code Generation failure loop-back (2.6.20) #616 loop-back, escalate to HITL on the same budget rules.
• Traceability — each finding links to the artifact(s) it changed; receipts bind to the amended source state; unaddressed findings stay visible as open items.
• No progress destruction — ingestion must not mark unrelated downstream stages not-started (the Jump anti-behavior).

Acceptance

• A 40+ item review file against a completed feature intent produces a bounded triage+touch plan in one pass; on approval only the implicated stages/files change.
• Findings with no clear target are surfaced for user disambiguation, not silently dropped.
• Re-running the same file after fixes reports remaining items only (idempotent on addressed findings).

Notes

• Depends on the amend scope + warm re-scan issues.
• Coordinate the review-comment ingestion path with PR-level review (feat(ci): add adversarial AI PR review agent #799). Roadmap Goal 6 (in-place enrichment).
• Author in core//harness/; add tests at the appropriate tiers.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of the project license.

[SaRedfiche]

Parking this — it depends on machinery we're no longer building for now.

Batch feedback-ingestion was specified to run under the amend scope (/aidlc amend --from-review <path> / --from-requirements <path>), triaging a batch of findings and driving them in place across the implicated stages of the intent already under way. Two of its three prerequisites have been withdrawn:

• The amend scope (feat: first-class amend scope for iterating on an existing intent (distinct from Jump/Redo/new-flow) #1123) is closed not-planned — as a scope route it collapsed into refactor (identical stage set), and the "operate on the active intent in place" behavior it needed lives in the dispatch/composer layer, not in scope membership.
• The composer auto-proposal / dispatch path that would bind an amendment to the running workflow was deliberately dropped from that effort.

Without those, there is no in-place amendment flow for ingestion to plug into, and the --from-review / --from-requirements flags have nothing to land on. Ingestion should be revisited only after (and on top of) a real dispatch-layer design for iterating on an active intent — at which point it's a natural follow-on. Reopen or re-file then.

Warm re-scan (#1122) is unaffected and shipped separately as PR #1125.