Skip to content

fix(gates 5, 23, 24): a comment must not spend a window, or invent a gap (#415 class, #423) - #440

Merged
rubenvdlinde merged 1 commit into
mainfrom
fix/gates-423-fp-runner
Aug 13, 2026
Merged

rubenvdlinde merged 1 commit into
mainfrom
fix/gates-423-fp-runner

Conversation

@rubenvdlinde

Copy link
Copy Markdown
Contributor

Closes part of #423 — the false-positive half of the #415 comment class, runner side. Companion PR for the checker side: gates 9, 27, 28, 46, 47, 61, 62.

gate-5 — a docblock was charged against an attribute's visibility

_head_block takes the union of {the contiguous annotation run, a 20-line slice}. For a multi-line attribute the run breaks at the closing line, so the slice is the only thing that can reach the opener — and an ordinary docblock between them spends it. Two fixtures differing in nothing but the length of that docblock, both carrying a real three-line #[AuthorizedAdminSetting(...)]:

 2 doc lines -> [gate-5] route-auth: PASS
30 doc lines -> [gate-5] route-auth: FAIL — 1 routed method(s) missing auth attribute

The attribute is present in both. A security gate telling an author that documenting the endpoint is what made it unsafe, whose cheapest remedy is to delete the paragraph.

The walk now steps over a multi-line attribute by structure — on a line that closes more brackets than it opens it scans up for the opener, requiring the brackets between to balance — rather than by distance. The same repair check_contract_coverage documents at its L208, "WHY THIS IS NOT A LINE WINDOW". It is bounded at 40 lines, must land on a real attribute opener at line start, and the previous-member clamp still cuts the window at the last }.

The anti-widening pair caught something. On multiline-attribute-borrow — the same file plus one unguarded routed method below it — the OLD runner reports FAIL — 2, flagging the genuinely guarded method as well. The new one reports FAIL — 1 and names the unguarded one. The change removes exactly the false finding and keeps the real one, in the same fixture.

gate-23 — the one grep-kind capability row never read code

A docblock recording that the class deliberately does not do the thing produced a finding identical to a real one: or_abstraction_findings=1, dropping to 0 once the candidate is re-asked on its code lines. A real statement still fires. The PDOK rule in the same file has routed through _code_lines since it was written; the capability rows never got it. grep -rln stays as the candidate finder.

gate-24 — a comment invented a coverage gap

Both probes deciding na vs structural were raw greps, so two comments describing registrations that were removed made the gate claim the repo has a server↔JS pair nobody correlated. Not a red gate — a fabricated gap in the arithmetic that decides whether a run's coverage is believed:

before  COVERAGE: 17 of 65 … (17 of 18 applicable gates ran)
after   COVERAGE: 17 of 65 … (17 of 17 applicable gates ran)

and gate-24 is the only verdict that moves in the whole 65-gate run. Real registrations stay structural.

Masks, not new strippers

gate-24 shells out to lib/source_scope.py (--mask php, --mask js-comments), the same way gate-20 already does; gate-23 reuses its own file's _code_lines. A file type with no mask is judged raw — the pre-#423 behaviour, erring towards structural, the stricter side — and .vue is deliberately in that set, because masking an SFC routes through the delimiter handling under repair in #424 and a hole there would let a comment remove a real registration from the count. Same for a mask that fails to run: the raw answer stands.

Arms

gate-5: two new fixtures, four assertions (suite 37 → 41). gate-23: two assertions. gate-24: two assertions (suite 42 → 44). With the runner and the linter restored from origin/main, exactly the evidence arms flip and the controls do not.

⚠️ gate-23's first arms were written as exit-code assertions and BOTH passed on the unfixed script. The capability rows have their own bake-in epoch, so the suite's forced BLOCK mode never reaches them and WARN mode exits 0 whatever it found — the arm was measuring the epoch, not the finding. They assert the file list now, and force the capability epoch as a second witness. Agreement between the two arms was the tell.

tests/run-helper-suites.sh: 83 passed, 0 failed, 2 quarantined (both pre-existing and documented).

🤖 Generated with Claude Code

…gap (#415 class, #423)

The runner-side half of #423. Same class as the seven checkers in the
companion change, three different mechanisms.

gate-5 — A DOCBLOCK WAS CHARGED AGAINST AN ATTRIBUTE'S VISIBILITY.

`_head_block` takes the union of {the contiguous annotation run, a 20-line
slice}. For a MULTI-LINE attribute the run breaks at the `)]` line — it
matches none of the five run patterns — so the SLICE is the only thing that
can reach the `#[`, and an ordinary docblock between them spends it. Measured
on two fixtures differing in NOTHING but the length of that docblock, both
carrying a real `#[AuthorizedAdminSetting(\n Application::APP_ID\n)]`:

     2 doc lines -> [gate-5] route-auth: PASS
    30 doc lines -> [gate-5] route-auth: FAIL — 1 routed method(s) missing
                      auth attribute

The attribute is present in both. This is a SECURITY gate telling an author
that documenting an endpoint is what made it unsafe, and the cheapest way to
clear it is to delete the paragraph.

The walk now steps over a multi-line attribute BY STRUCTURE — on a line that
closes more brackets than it opens it scans up for the `#[` that opens them,
requiring the brackets between to balance — rather than by distance. Same
repair `check_contract_coverage` documents at its L208, "WHY THIS IS NOT A
LINE WINDOW". It cannot run away: the scan is bounded at 40 lines, it must
land on a real `#[` at line start, and the previous-member clamp still cuts
the window at the last `}`.

The anti-widening pair is the point, and it caught something. On
`multiline-attribute-borrow` — the same file plus one unguarded routed method
below it — the OLD runner reports FAIL — 2, flagging the genuinely guarded
`save` as well as `open`. The new one reports FAIL — 1 and names `open`. So
the change removes exactly the false finding and keeps the real one, in the
same fixture.

gate-23 — THE ONE `grep`-KIND CAPABILITY ROW NEVER READ CODE.

A docblock saying *"We deliberately do NOT set a Postgres search_path here —
isolation is OpenRegister's job (ADR-022)"* produced
`[or-capability:tenant-boundary]` against a file that does exactly what the
ADR asks: 1 finding, dropping to 0 once the candidate is re-asked on its code
lines. A real `SET search_path` still fires. The PDOK rule in the same file
has routed through `_code_lines` since it was written; the capability rows
never got it. `grep -rln` stays as the candidate finder.

gate-24 — A COMMENT INVENTED A COVERAGE GAP.

Both probes deciding `na` vs `structural` were raw greps, so

    // TODO: one day we could do `new LeafDescriptor(...)` here
    // We used to call registerIntegration({...}); removed in 2.4

made the gate claim the repo has a server↔JS pair nobody correlated. Not a
red gate — a fabricated gap in the arithmetic that decides whether a run's
coverage is believed. Measured on the fixture, the runner's own summary:

    before  COVERAGE: 17 of 65 … (17 of 18 applicable gates ran)
    after   COVERAGE: 17 of 65 … (17 of 17 applicable gates ran)

and gate-24 is the ONLY verdict that moves in the whole 65-gate run. A real
`new LeafDescriptor(` plus a real `registerIntegration(` stays structural.

MASKS, NOT NEW STRIPPERS. gate-24 shells out to lib/source_scope.py
(`--mask php`, `--mask js-comments`), the same way gate-20 already does;
gate-23 reuses its own file's `_code_lines`. A file type with no mask is
judged RAW, which is the pre-#423 behaviour and errs towards `structural`,
the stricter side — and `.vue` is deliberately in that set, because masking
an SFC routes through the `<!-- … -->` handling #424 is repairing and a hole
there would let a comment REMOVE a real registration from the count. Same for
a mask that fails to run: the raw answer stands.

ARMS. gate-5: two new fixtures and four assertions in test_gate_route_auth.sh
(37 -> 41). gate-23: two in test_or_abstraction_pdok.sh. gate-24: two in
test_gates_23_33_never_green_over_nothing.sh (42 -> 44). With the runner and
the linter restored from origin/main, exactly the evidence arms flip —
gate-5's "29-line docblock" and "exactly ONE finding", gate-23's "silent:
docblock", both of gate-24's — and the controls do not.

⚠️ gate-23's first arms were written as `assert_rc` and BOTH passed on the
UNFIXED script. The capability rows have their own bake-in epoch, so the
suite's forced BLOCK mode never reaches them and WARN mode exits 0 whatever
it found: the arm was measuring the epoch, not the finding. They assert the
FILE LIST now, and force the capability epoch as a second witness. Agreement
between the two arms was the tell.
@rubenvdlinde
rubenvdlinde merged commit 8387110 into main Aug 13, 2026
34 checks passed
@rubenvdlinde
rubenvdlinde deleted the fix/gates-423-fp-runner branch August 13, 2026 08:53
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants