fix-forward: reduce drip-feed review rounds for implementer and responder

[microsasa]

Problem

Copilot code review drip-feeds ~2 comments per pass. When the responder fixes only the flagged instances, Copilot re-reviews and finds more of the same class in the next round. The implementer also ships code with patterns the reviewer predictably catches (stale docstrings, weak assertions, dead references).

Evidence

Analysis of 32 pipeline PRs:

• 9 PRs (28%) had the same issue class repeated across review rounds
• Fix-forward would have saved ~12 review rounds across those 9 PRs
• In 4 additional PRs (fix: remove dead alias, clean report.__all__, and fix Pydantic default_factory (#776) #784, perf: short-circuit TOOL_EXECUTION_COMPLETE in _first_pass (#772) #778, fix(parser): eliminate O(N) failing stat() calls for absent plan.md (#763) #767, fix: remove dead-code parse_data() and EventData from public API (#670) #674), the reviewer caught 2 instances of the same class in round 1 that the implementer could have caught before pushing

Top repeated issue classes

Class	PRs affected	Agent
Stale docstrings/comments	#736, #781, #719, #706, #768	Both
Weak/misleading test assertions	#654, #726, #760	Both
Same bug in parallel code paths	#736, #727	Responder
Dead code references	#699	Both
Naming mismatches	#654, #719	Implementer

Solution

Responder: fix-forward scan (step 5)

After addressing each review comment, the responder identifies the class of issue and scans ALL changed files for other instances. Prevents the reviewer from flagging the same class in the next round.

Implementer: pre-push self-review

Before committing, the implementer audits its own changes for the reviewer's most common findings:

• Docstring accuracy (does it still match new behavior?)
• Test assertion strength (exact match vs substring)
• Dead references (renamed/removed → grep for stale refs)
• Parallel code paths (same bug/gap in sibling functions)
• Naming consistency (test names match behavior)

Investigation findings

See #782 for the full research into Copilot review behavior (comment limits, configuration options, platform changes, workarounds).

Key facts:

• No configurable comment limit exists — hardcoded in the product
• GitHub's blog warns against "be more thorough" instructions — adds noise
• Average 5.1 comments/review, 29% produce zero
• Instructions have 4,000 char limit for code review
• May 2025: "80% more comments per PR" (platform change, not configurable)
• Jul 2025: Lifted file count limits for large PRs