[adithya-s-k]

Arc 2 — pr_runtime (work in progress)

Per plan §4 Arc 2. Stacked on #31 (Arc 1 pr_diff) which is stacked on #30 (sweep harness).

This PR lands the optimization + tests; the full 38-repo sweep is deferred until cost approval. See the findings doc for full context.

Summary

Optimization — opt-in F2P filter relaxation

Plan §4 Arc 2 known issue: ~17-27% of candidates drop on no_fail_to_pass because the PR's test patch added or modified tests that already passed before the fix landed. These are real production bug fixes whose test coverage doesn't follow the strict SWE-bench shape.

• Add PRRuntimeOptions.allow_no_f2p_with_test_patch: bool = False.
  When set via --pipeline-opt allow_no_f2p_with_test_patch=true, the pipeline accepts F2P=0 candidates iff the test_patch is non-empty AND the P2P set has ≥1 entry.
• Decision lives in a pure helper _should_skip_no_f2p for testability.
• 6 new unit tests in tests/test_pipeline_pr_runtime.py covering the strict path + each relaxation guard.

Harness bugfix — discovered by the smoke

Smoke surfaced that `repo2rlenv generate` exits 1 when emitted=0 (cli.py:275). For pr_runtime that's a normal filter outcome — the sweep was misreading it as a cell failure.

• scripts/v083/sweep.py now treats "exit non-zero + out_dir exists + no task dirs" as step=done, candidates=0. True crashes (out_dir missing) still land as failed.
• Also wires `--pipeline-opt KEY=VALUE` through the sweep CLI so arc-specific options flow without code edits.
• 3 new unit tests for the option parser; suite at 658 passing.

Smoke evidence (limited)

Sweep	Repos	envs/cell	Outcome
#1	pallets/click	2	sweep reported step=failed because of the exit-code issue; pipeline itself worked. Surfaced the harness bug.
#2	pallets/click	3	3 PRs: 2 no_test_patch, 1 no_fail_to_pass (P2P=0 → relaxation correctly didn't rescue)
#3	pallets/click + pallets/werkzeug	5 each	2/2 cells step=done, 0 emitted, $0 cost. Harness fix verified.

The 10 most-recent PRs in click+werkzeug happen to not include any matching the relaxation pattern. To exercise the rescue path requires historical PRs (--pipeline-opt since=2024-01-01) or higher-activity repos like pydantic/pydantic or psf/requests — deferred to the full sweep.

What's pending for full Arc 2 completion

1. Full 38-repo sweep — should yield ~75-85 verified envs. Cost envelope: ~$80-250 (bootstrap LLM + T2 Haiku + optional T4 Sonnet). Wall time: 6-12 h.
2. HF push to AdithyaSK/repo2rlenv-v083-pr_runtime.
3. Findings update with concrete T3 yield + relaxation-impact delta.
4. Per-language log-parser polish for Tier C — deferred until we see real Tier C failure modes from the full sweep.

Test plan

• uv run pytest -q — 658 passing, 2 skipped (+6 new F2P-gate tests, +3 new option-parser tests)
• uv run ruff check . + uv run ruff format --check . — clean
• Smoke: sweep.py --pipeline pr_runtime --repos pallets/click pallets/werkzeug --pipeline-opt allow_no_f2p_with_test_patch=true → 2/2 cells step=done
• Full 38-repo sweep — pending user OK on cost
• HF push — pending sweep
• Consumer-side harbor run --agent oracle smoke on a random task — pending sweep

Out of scope

• LLM-judged QA gate (deferred to v0.9 per plan §11).
• Per-PR test-cmd normalization edge cases beyond what fix: bootstrap saves executable test_cmds; normalize strips -q #23 already shipped.
• T4 Sonnet agent eval — covered separately once T3 yield lands.

Notes for reviewer

• Stacked on v083 Arc 1: pr_diff sweep + info-leak hardening (127 envs) #31 → v083: sweep harness (scripts/v083 + tests) #30. When upstream PRs merge, GitHub auto-rebases this base.
• No pyproject.toml bump (stays at 0.8.2.post3 per plan §5).
• No Co-Authored-By trailer.

[adithya-s-k]

Closing as part of v0.8.3 cleanup — restructuring to one-PR-at-a-time per request. Work preserved on the branch; see PR #31 for the active cleaned-up arc.