fix(bp-155): decide the L0a oversize cut over the canonical body (A1.2)

[ascalva]

Summary

Completes bp-155 — the L0a oversize cut in _l0a_chunks (core/ingest/code_corpus.py) now
decides the whole↔windowed split over the canonical body (len(body) <= max_chars),
not the header-bearing full text (len(header + body) <= max_chars). This closes issue #31,
bp-151's parked residue, per Amendment A1.2 of dn-vector-membership-store (owner ruling
2026-08-06). This completes D0: bp-151 took the aggregate rename cost from 11,096 → 3;
this takes it to 0.

• core/ingest/code_corpus.py — the one-token change (:152), and the KNOWN RESIDUE comment
  replaced with one recording the decision is now canonical-body-scoped, per A1.2. Nothing else
  in the block moved — text=full stays text=full, so L0a embed text keeps its coordinate
  header (D0/R7 untouched).
• tests/unit/test_code_corpus.py — bp-151's deliberate tripwire,
  test_l0a_oversize_threshold_is_the_one_rename_residue, converted to
  test_l0a_oversize_cut_is_canonical_body_scoped: same straddle-precondition fixture (the
  rename still crosses the OLD header-bearing threshold), the mint-1 assertion flipped to
  mint-0, docstring records it was bp-151's tripwire and that it reddened exactly as designed.

closes #31

Measured (real chunkers, real files, not a fixture — 2026-08-06)

Item 2 — aggregate rename cost: 0. Over all 580 tracked .py files: for each file,
derive L0a/L0b/L1 at its own path and at a length-changed moved path (same directory, longer
basename), sum new (layer, content_hash) atoms across all three layers.

• Post-fix: 0 new atoms, 0 files affected.
• Pre-fix (same script against the stashed-out old code, on this tree): 23 new atoms
  across 23 files — nonzero, confirming the measurement has teeth. (Larger than bp-151's
  original 3 because the tree has grown since 45c4a15; not a discrepancy.)
• A1.5's first falsifier — a residue meaning a fourth path-dependent site — does not fire.
  No issue filed.

Item 3 — L0a boundary-change census: 72 groups across 57 files, all in-band. Comparing
L0a (path, qualname) groups' canonical-body hashes pre/post fix: 73 groups changed
across 58 files; 72/57 are rule-driven, and 1
(core/ingest/code_corpus.py::_l0a_chunks itself) is a measurement artifact — this PR's own
comment edit inside that function's body, not a rule effect. Verified all 72 genuine
changes fall inside the predicted affected band
(max_chars - len(header) <= len(body) <= max_chars) — A1.5's second falsifier (a change
outside the band) does not fire. (Was ≈123/95 at 45c4a15; the tree has moved — this is
what it reads now, not a regression from that number.)

A side-finding, not a defect. For groups in the affected band, the OLD rule routed the
body through chunk_text (core/kernel/ingest/chunk.py), whose _blocks calls .strip()
on a block even when the whole body is a single block under budget — silently stripping the
first line's leading indentation for nested symbols (a class method's body starts
indented). E.g. config/secrets_backend.py::VaultClient.mint_token: old rendering was
'def mint_token(self, role...' (dedented, corrupted); the raw body is
' def mint_token(self, role...' (correct). The new rule takes the whole branch directly
for all 72 slices, so their canonical bodies (and embed text) are the exact, unmangled source
for the first time — a side benefit, not new behavior to chase, and not filed as a separate
issue.

Gate — exact results

Each leg run separately, never piped (pytest -q | tail returns tail's exit code).

$ uv run ruff check .
All checks passed!
(exit 0 — one E501 introduced by the converted test's trailing comment was fixed before this run)

$ uv run mypy core agents eval ops scheduler scripts
Success: no issues found in 262 source files
(exit 0)

$ uv run mypy
...
tests/integration/test_cron.py:91: error: Argument 2 to "cron_handlers" has incompatible type "FakeCurator"; expected "Curator"  [arg-type]
Found 69 errors in 20 files (checked 563 source files)
(exit 1, as designed — tail count 69, unmoved from the pinned baseline)

$ uv run python -m ops.type_gate
Tier-2 membership: OK — every core-importing top-level package is in [tool.mypy].files
Bare-ignore scan: OK — every `# type: ignore` in the checked region carries an error code
Raw shimmed-dependency imports (type-gate) REPORTED (parked, non-fatal) — one pre-existing,
unrelated waiver (finding-0223)
(exit 0)

$ uv run pytest -q
...
FAILED tests/e2e/test_dream_v2_live.py::test_dream_v2_synthesizes_grounded_themes_live
FAILED tests/integration/test_worktree_enforcement.py::test_a_deny_cross_worktree
FAILED tests/integration/test_worktree_enforcement.py::test_c_unsafe_direction_narrow_not_loosened
FAILED tests/integration/test_worktree_enforcement.py::test_d_no_pointer_is_no_plan_not_main_fallback
FAILED tests/unit/test_core_self_containment.py::test_core_imports_nothing_outside_core
5 failed, 2427 passed, 15 skipped, 12 warnings in 224.14s (0:03:44)

5 failed / 2427 passed / 15 skipped — exactly the three known-red classes local-only, green
in CI: the finding-0103 core-self-containment ratchet (1), tests/e2e/test_dream_v2_live.py (1),
and tests/integration/test_worktree_enforcement.py (issue #13 / finding-0280) (3). Same test
names, same count bp-151 reported for its own build. tests/e2e/test_scheduler_live.py's known
flake did not fire in either run below.

Diff-innocence proven: stashed both changed files back to a byte-identical origin/main
(git diff --stat origin/main HEAD empty), re-ran the full suite on the clean base — same 5
failures, same names, 5 failed, 2427 passed, 15 skipped in 249.76s.

Test plan

• ruff check . clean
• mypy core agents eval ops scheduler scripts — 0 issues, 262 files
• mypy (argless) — tail 69, unmoved
• ops.type_gate — OK
• pytest -q — 5 known-red / 2427 passed / 15 skipped, proven diff-innocent against clean
  origin/main
• test_l0a_oversize_cut_is_canonical_body_scoped passes; inverted (stash the fix) to
  confirm it reproduces the original mint-1 residue, then restored
• Aggregate rename cost measured at 0 over all 580 tracked .py files (not a fixture)
• Boundary-change census (72/57) verified entirely inside the predicted affected band

🤖 Generated with Claude Code

https://claude.ai/code/session_011LZZQPyGsoeGL73cbbEp3U