perf+fix: faster feature builds — deterministic git fast-paths, scope-proportional planning, LLM shape normalization

[AbirAbbas]

Summary

One PR for the full improvement round that came out of a live full-pipeline benchmark (1h43m for a 2-file task on openrouter/deepseek/deepseek-v4-pro, 88 agent calls, plus a second run that crashed at 48m). Three commits, three concerns:

1. fix: normalize LLM-emitted scalar shapes on issue data — the 48m run died with 6 DAGState validation errors on checkpoint reload: the replanner emitted acceptance_criteria as a plain string, assignment bypassed validation (validate_assignment off), and the poisoned checkpoint failed re-validation. Now every LLM ingestion boundary (fix generator, replanner update/add paths) coerces str → [str] (ensure_str_list / normalize_issue_dict; Go StrList with tolerant UnmarshalJSON), and the schemas carry mode="before" validators so even pre-fix poisoned checkpoints still load. The exact incident is a regression test.
2. perf: deterministic git fast-paths — LLM-driven git bookkeeping was 23 of 88 calls (7 workspace setups, 7 mergers, 9 cleanups), ~20 minutes. Worktree creation and cleanup are always mechanical; merges are mechanical when conflict-free. They now run as plain git (git_fast_path.py / internal/dag/gitfast.go): the merger agent only ever sees branches that actually conflict, and single-branch levels no longer request integration tests (per-issue tests already covered them). deterministic_git=true by default; false restores the agent-driven path, and any fast-path failure falls back to the agent — behavior never gets worse. Branch/worktree naming matches prompts/workspace.py exactly, so fast-path and agent-created state stay interchangeable (resume, mixed fallbacks, cleanup).
3. perf(prompts): scope-proportional decomposition — the sprint planner turned the 2-file task into 7 issues with 12 acceptance criteria, and the fix generator's "exactly ONE failed criterion per fix issue" rule turned one verification failure into 7 fix issues (each a full coder+review cycle). The sprint planner now carries an explicit sizing rubric (a 1–3 file change is ONE issue; calibration table; 3–5 ACs per issue), and the fix generator groups failures sharing a root cause or files into one fix issue. Go prompt mirrors updated; golden fixtures regenerated from the Python truth.

Test plan

• tests/test_llm_shape_normalization.py + Go strlist_test.go / normalize_llm_shapes_test.go — coercion at every ingestion boundary, tolerant re-validation, and a regression test reproducing the exact crashed-checkpoint reload
• tests/test_git_fast_path.py + go/internal/dag/gitfast_test.go — naming parity with the agent convention, clean-merge/conflict-abort semantics (integration branch left clean), conflict-only agent handoff, resume reuse, cleanup, flag-off and fallback paths
• Existing executor suites unchanged and green (fallback keeps scripted-agent tests working)
• Full suite at CI parity (no .env): 1079 passed, 1 skipped + compileall clean; Go gofmt/build/vet/-race with GOWORK=off
• Live before/after benchmark on the same task/model: 1h43m39s / 88 calls / 7 issues → 41m18s / 34 calls / 2 issues, verification passed all 10 ACs, zero git-agent calls — full numbers in the comment below

---

🤖 Generated with Claude Code

[AbirAbbas]

Live before/after benchmark

Same task (add a retry decorator + tests to a small Python repo), same model (openrouter/deepseek/deepseek-v4-pro, open_code runtime), same config, fresh repo copy each run:

	Before (main)	After (#102 + #103)
Wall clock	1h 43m 39s	41m 18s (2.5×)
Agent calls	88	34
Issues planned	7 (12 ACs)	2
Git agent calls (setup/merge/cleanup)	23	0 — all deterministic
Replanner churn	1 full replan (~33 min)	none
API cost	≈$4.50	≈$1.31
Outcome	6/7 issues, 5 tests	2/2 issues, verification passed all 10 ACs, 6 tests

Evidence the fast paths engaged: the delivered branch's history contains the fast path's merge(level-N): <branch> commits, and the reasoner timeline has zero run_workspace_setup / run_merger / run_workspace_cleanup calls. A separate pre-fix run of main also crash-failed at 48m on the exact bug #102 fixes (internal trace: exec_20260720_110500_m36lxxcq), so the before-side is arguably worse than the table shows.

Remaining time is planning (~10m, unchanged) and coder/review iterations (model-bound) — the structural overhead this PR targets is gone.