feat: intake reconcile --repo — read the target repo for staleness evidence

[Jammy2211]

Summary

Leg 3 of the draft-staleness work. Closes #223.

Legs 1 and 2 shipped Mind-local signals — the PyAutoMind lifecycle.py gate keys, and the intake reconcile re-rank that cut the flag rate from 96/148 to 29/135. Both are structurally blind to two of the five findings the 2026-08-09 draft/ sweep confirmed: one whose evidence sat in a sibling prompt rather than a completion record, and one whose fix shipped upstream with no completion record written at all. The parent prompt measured that no amount of re-ranking reaches those — every attempt cost precision without gaining truth. Only reading the target repo can. This adds that read.

intake reconcile --repo TARGET resolves the target through PyAutoMind repos.yaml plus the sizing faculty's existing repo_aliases, makes a cached shallow clone, greps for the identifiers each prompt names, and reports them as file:line evidence in a needs-review band. The resolved sha is printed so a verdict is re-checkable.

The trap it is built around

Upstream hits deliberately do not feed overlap_score. draft/bug/autofit/test_mode_bypass_ordered_assertion_ties.md names five identifiers — FitException, check_assertions, ignore_assertions, instance_for_arguments, instance_from_vector — and all five are on PyAutoFit main. The prompt is confirmed NOT shipped: main catches exc.FitException in the TEST_MODE bypass, which looks exactly like the requested fix, but the catch wraps only the likelihood call while model.instance_from_vector (where check_assertions raises on an ordering tie) sits on the line before the try.

Scoring presence as evidence would rank that prompt high and make the one mis-grade this tool must never make. Verified end-to-end against PyAutoFit fbe9f45d:

        medium  overlap=8.0  upstream=7.5  ep_hierarchical_scale_collapse_moment_match.md
  needs-review  overlap=0.0  upstream=6.0  test_mode_bypass_ordered_assertion_ties.md
  needs-review  overlap=0.0  upstream=4.5  plot_functions_discard_kwargs.md

Surfaced, with its evidence, at overlap_score 0.0 — never called shipped. That is structural, not a tuned threshold: no number of upstream hits can reach a Mind-local band.

API Changes

Additive only; no existing signature or output changes for callers that don't pass the new flag.

• intake reconcile gains --repo TARGET (opt-in; default path byte-for-byte unchanged).
• reconcile() gains source_reader= and upstream_meta= keyword arguments, both defaulting to the previous behaviour.
• New public helpers: resolve_repo(), upstream_reader().
• New confidence band needs-review, emitted only under --repo.
• Result dict gains upstream (slug/sha/date) and per-suspect upstream_score.

Test Plan

• pytest tests/test_intake_reconcile_ranking.py — 16 passed (8 pre-existing + 8 new)
• pytest tests/ --ignore=tests/test_skill_install.py — 330 passed
• Default path unchanged: still 29 suspects of 135 scanned, 7 high / 22 medium
• Live run against PyAutoFit fbe9f45d — trap lands in needs-review, not shipped
• --repo workspaces|priors|health_fixes|graphical_ep → exit 5, naming real candidates
• Alias resolution: autofit / PyAutoFit / pyautofit → PyAutoLabs/PyAutoFit
• bin/check_skill_line_counts.sh — within the 200-line mandatory-load budget

Ship gate: RED, on a pre-existing failure unrelated to this diff

PyAutoHeart is not reachable from this environment, so the documented fallback gate (pytest -x, any failure = RED) applied and came back RED. Recording it explicitly rather than quietly passing:

• Reason: assert "SKIP intake (Codex skill" in result.stdout — tests/test_skill_install.py, 2 failures.
• Not caused by this diff. Both reproduce at merge-base a50efc3 with the diff entirely absent. The assertion is about the installer symlinking skill directories into a Codex home, not about the skills/intake/intake.md edit here. bin/install.sh takes its web-github / ci-only branch (:282) because $HOME/Code/PyAutoLabs does not exist, so it never performs the work the test asserts on.
• A human acknowledged the RED reason and authorized opening this PR on that basis.

Design notes — two decisions worth reviewing

Noise filter: two named classes, not a spread threshold

The first real run matched TypeError (37 files in PyAutoFit) and autofit_workspace (26) — builtins and repo names, pure noise. Filtering by upstream file-spread was tried and rejected: the counts do not separate, because instance_from_vector is a real signal at 22 files, just under autofit_workspace at 26. Any threshold dropping the noise also drops one of the trap's own identifiers. Filtered the two classes explicitly instead; one all-noise prompt dropped out of the results.

Plain shallow clone, not treeless

The parent prompt suggested --depth 1 --filter=blob:none. That is wrong here — this leg greps source, and a treeless clone refetches every blob on demand to answer that. Plain --depth 1 with GIT_LFS_SKIP_SMUDGE=1, cached under $PYAUTO_BRAIN_CACHE (default ~/.pyauto-brain/upstream). Full run against PyAutoFit takes ~3.6s warm.

First network access in PyAutoBrain

Nothing under agents/ previously used urllib, requests, gh or git clone. This is strictly opt-in: without --repo no clone, socket or subprocess runs, and test_default_path_makes_no_network_access monkeypatches both socket.socket and subprocess.run to detonate if that ever changes. AUTONOMY.md records the surface and states that widening it is a new decision.

Multi-repo targets exit 5 rather than guessing — workspaces (23 prompts across four work-types), health_fixes, priors and graphical_ep are among the largest buckets in draft/, so a guess would be confidently wrong at scale.

Tests — 8 new, all hermetic

Nothing under tests/ clones — the upstream read is injected through the source_reader seam, preserving the file's existing property that every fixture is a fictional Mind tree in tmp_path.

• test_upstream_presence_never_produces_a_shipped_verdict — the trap, in miniature
• test_upstream_evidence_cannot_inflate_a_mind_local_band — the structural defence behind it
• test_one_upstream_identifier_is_not_a_signal
• test_builtins_and_repo_names_are_not_upstream_evidence
• test_default_path_makes_no_network_access
• test_upstream_mode_still_never_writes
• test_a_multi_repo_target_is_refused_not_guessed
• test_a_real_repo_target_resolves_to_its_slug

Generated by the PyAutoLabs agent workflow.