Do not report a required subprogram verified when its gate left it ungated - #35
Merged
Conversation
…gated
verify_recipe_candidates decided a required subprogram's gates_passed from
the owner unit's gate statuses alone. A differential gate whose oracle has
no reference for a subprogram lists it as ungated (metrics["ungated"],
{name: why}) and passes the unit on the rest, so a required subprogram that
nothing ever compared came out observed/gates_passed/accepted, the report
accepted, and the training loop merged the candidate.
Read each required gate's verdict for its ungated names and treat a
required subprogram ungated on any owner unit as not gates_passed. The
report then rejects with subprogram.required_coverage, the code the
campaign already routes to repair. Unrequired subprograms keep today's
behaviour: recorded as unverified, not blocking.
Found by the lab: 39f9f8d stubbed LAPACK externals in the oracle build and
listed the subprograms reaching them as ungated; splines' required spline3
and spline3pars passed without being compared.
Signed-off-by: 2st <mouseisnotalwaysmouse@outlook.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
chenyueqi
added
bug
chenyueqi
added a commit
that referenced
this pull request
Sep 6, 2026
… and #35 The table is unchanged. Per-subprogram metrics now carry PR #31's `shaped` and `input_profile` and PR #33's `declined` by reason, and minpack's chkder is named for its declined draws as PR #33's own re-record named it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018vrAgr6x1yLT47EumS39L5
chenyueqi
added a commit
that referenced
this pull request
Sep 6, 2026
…-bounds The CLUBB stack and this branch extended the same functions in different directions; where git could not merge them, this is what was decided: - An Expr has both a `str` node (a character value) and a `call` node (a folded intrinsic); `render` spells both. Constant folding is main's typed `with_integer_division`, with the `str` short-cut kept. - One ALLOCATE scanner walks parentheses across continuation lines for both readers: an object's components (`COMPONENT_ALLOCATION`) and a module allocatable's own shape (`ALLOCATION`); `ALLOCATE_THIS` goes. - The recorder keeps this branch's guard for an array the run never allocated -- poison of the planned shape -- and generalizes it to CLUBB's case: an extent the run owns (`extent_args`, written from `size()`) is spelled zero under the guard, so `c%coef` under `sclr_dim = 0` is recorded as `(ngrdcol, 0)` of nothing, the shape the kernel's dims declare, rather than main's `(0,0)` record. `_record_component` goes; the CLUBB test asserts the new spelling. - The replay reads a logical (`T`/`F`) before it takes any other text as a character value. The flat adapter names a type only where a dummy of it is declared, and reaches a private specific through its generic. - The JAX call rewrite takes main's host-call preamble and positional list, narrowed to the emitted convention: optionals as keywords, OUT scalars omitted, an OUT array passed only where it is the caller's buffer (the `buffer` mark, not `dims`). The elided-buffer table is main's positional one with this branch's temporary for an OUT scalar the anchor unpacks; an unpassed OUT array stays refused. - The single-exit rewrite is this branch's flag-based one (returns inside loops, the guarded remainder as one block), with main's placement of the flag and placeholder after the statements before the first return, the placeholder shaped like the final value where that is already bound. The flag guard now applies at every block level through the visit of each branch, which is what main's `_fold_returns` covered (`if (bad) then; if (worse) return; y = y + 1; return`); that helper and `_guard_after_returns` go. - The kernel lowerer knows both the trace-time statics and the names that size an array or bound a range; a literal store to either stays Python. - The JAX runtime exports both sides' shims once; SUM folds sequentially along an axis too. The NumPy SUM is main's for the same reason. - `_ret`-flag tests: the statics marker may precede a loop's done flag. - Ruff and mypy on this branch's own files: long lines wrapped, a re-exported constant imported from its home, duplicate runtime definitions dropped. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_018vrAgr6x1yLT47EumS39L5
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this is
verify_recipe_candidatesdecided a required subprogram'sgates_passedfrom the owner unit's gate statuses alone. A differential gate whose oracle has no reference for a subprogram lists it as ungated (metrics["ungated"],{name: why}) and passes the unit on the rest, so a required subprogram that nothing ever compared came out observed / gates_passed / accepted, the report was accepted, and the training loop merged the candidate.gates_passed. The report then rejects withsubprogram.required_coverage, the code the campaign already routes to repair. Unrequired subprograms keep today's behaviour: recorded as unverified, not blocking. Tests intests/test_phases.py.Found by the training lab: an agent repair (
39f9f8dthere) stubbed LAPACK externals in the oracle build and listed the subprograms reaching them as ungated;splines' requiredspline3andspline3parspassed without ever being compared. This is the upstream version of the lab's fix, rebased ontomain(1044bc4). The lab's agent-authored engine repairs that followed (C-keyword dummies as f2py names, negative-subscript redraws,LOGICAL INOUTscalars) are pending human audit and are not in this PR.Checks run locally (uv, Python 3.11, gfortran 13)
ruff check .clean;mypyclean (109 files);tools/check_signoff.py 1044bc4..HEADclean.pytestwithdev fortran translate verify: 891 passed, 20 skipped.tools/check_hygiene.py .: the same 4 lines asmain(docs/roadmap.md,src/recast/fortran/constants.py), nothing from this branch.ruff format --check .: the same 2 Markdown files asmain, untouched here.🤖 Generated with Claude Code