porch: consult raw outputs don't survive retries or builder cleanup — archive per-attempt artifacts

[waleedkadous]

porch: consult raw outputs don't survive retries or builder cleanup — archive per-attempt artifacts

Problem

Consult raw review outputs are ephemeral twice over:

1. Retries overwrite. Porch keeps one output file per iteration (<id>-<phase>-iterN-<model>.txt); a retry of the same iteration overwrites the previous attempt's output at the same path. Attempt-level artifacts — the primary debugging evidence when a lane misbehaves — are destroyed by the very retry that works around the misbehavior.
2. Cleanup deletes. Consult outputs live uncommitted in the builder worktree; afx cleanup after merge deletes them all.

Field impact (adopting workspace, details scrubbed)

Three incidents of the gemini/agy impl-lane flake (hollow skip stubs with misleading self-diagnosis — see #1166) were reported across two builders and two days. Raw evidence survived for exactly one of the three, and only because a builder manually copied the logs out of the worktree before they were overwritten/cleaned. The other two incidents are unverifiable — including the first one observed (an empty-sandbox false REQUEST_CHANGES, the #1051 class), whose worktree was cleaned after merge.

Debugging intermittent consult-tooling failures (#1051, #1166, #1032/#1033 family) is evidence-starved by design today.

Proposal

• Porch archives each consult attempt's raw output (and the stub, if the attempt was skipped) to a durable per-project location — e.g. codev/projects/<id>/consults/<phase>-iterN-attemptM-<model>.txt — never overwritten by retries.
• The archive dir is committed with the project state at the pr-ready gate (consistent with the "completion stats in status.yaml" convention), or at minimum survives in the main checkout after merge.
• Retention can be bounded (e.g. keep all attempts for the current iteration + final output per past iteration) if size is a concern; full impl prompts are ~1MB per attempt, so a cap or gitignore-with-local-retention split may be preferable to unconditional commit.

Non-goals

Fixing the flakes themselves (tracked in #1051/#1166 etc.) — this issue is about not destroying the evidence needed to fix them.

Areas: area/porch.