Repository navigation
gen/c: 32 of 43 tracked C copies are stale since 2026-07-05, and L2 does not check a spec changed without its copy #7103
Description
Problem
gen/ holds the generated copies of the specs (L2). On master 96cf7c8, 32 of the 43 tracked gen/c/*.c files that have a spec at specs/<path>.t27 are not what t27c gen-c of that spec prints today.
- Every one of the 32 was last written by one commit on 2026-07-05 (ea15cd5):
ar/*(7)base/ops,base/typescompiler/parserfpga/*(5)isa/registersmath/constants,math/sacred_physicsnn/attention,nn/hslmnumeric/*(10)queen/lotusvsa/ops
- They differ by 166 to 2626 diff lines each.
- The 11 that match are exactly the ones a tool or a gate compiles:
policy/own_languagequeen/priorityandqueen/review_valvetri/catalog/healthtri/t27b/*(5)ci/affectedandci/delivery_metrics
- Nothing in
scripts/,tools/,cli/or.github/reads the 32 stale copies. - Docs still point at them:
docs/NUMERIC_FORMATS_SSOT.mdlinksgen/c/numeric/gf16.cas the C codegen;README.mdnamesgen/c/numeric/as a header source.
A second gap lets this keep happening. tools/l2_regen_check.py (#6226) checks a modified gen/ file against t27c gen-<backend>. It does not check a modified spec whose tracked copy stayed as it was.
PR #7091 hit this. It added test asserts to specs/queen/priority.t27 and review_valve.t27 and did not regenerate their C copies, and no gate said so. tri priority builds gen/c/queen/priority.c. I found it by hand and fixed it in 5b338c7.
Measurement
Run from a worktree, so that use resolves; a spec copied to /tmp changes the output:
for g in $(git ls-files gen/c); do s=specs/${g#gen/c/}; s=${s%.c}.t27; [ -f $s ] || continue
t27c gen-c $s | diff -q - $g >/dev/null && echo SAME $g || echo DRIFT $g; done
gen/c/vsa/core.c has no spec at specs/vsa/core.t27 and is not counted.
Fix
- Gate, both directions. L2 also lists the specs a PR modified. For each one that has a tracked
gen/<backend>/<path>copy, it runs the same byte compare. A spec change with a stale copy then fails the way a hand-edited copy does. - The 32 stale copies. For each, either regenerate it with
t27c gen-c, which the only-t27 rule allows as t27c output, or delete it if nothing should read a C copy of that spec.- The docs that link a deleted copy are fixed in the same PR.
- gf16 is in gf16: bring conformance vectors, ring-088, property template, docs and gen/ in line with the #6940 decisions #6996's scope, so it is left to that issue.
Acceptance criteria
- The loop above prints no DRIFT line on master.
- A PR that changes
specs/queen/priority.t27without regeneratinggen/c/queen/priority.cfails L2.- Negative control: the same PR with the copy regenerated passes.
docs/NUMERIC_FORMATS_SSOT.mdandREADME.mdname only files that exist and match.
Found in the queen-actors loop (epic #6971) while landing #7077. Refs #6226, #6996.
Activity
gHashTag commentedon Oct 6, 2026
Scope correction and the fix, measured 2026-10-06T21:05Z on branch
claude/gen-c-drift-7103(local, three commits on master 96cf7c8; the PR opens after #7091 and #7092 land, one PR at a time).The issue undercounted the drift. It said 32 of 43 gen/c copies. Measured with every tracked copy regenerated by the master t27c, the drift was:
gen/rust/memory/notebooklm.rs;gen/c/vsa/core.candgen/verilog/vsa/core.v. Their specspecs/vsa/core.t27does not exist;specs/vsa_core.t27does.Decision: delete, not regenerate. Commit 2b74dc600 drops 65 tracked copies.
git grepof each deleted path, outside gen/ and docs/now). Prose mentions three paths: STATUS.md and FORMAT_REGISTRY.md say TF3 is generated togen/verilog/numeric/tf3.v, which stays true oft27c gen-verilog;fpga/uart.vandfpga/mac.vare named in an ADR, docs/NOW.md and a wave report, all history..gitignore.lexer::TokenKind::RBrace) is still refused -- 477 in the specs #5712 family; regeneratedgf4.creferencessacred_physics::PHI_INVverbatim).Kept: 14 copies. 11 C copies that match and are read (ci/affected, ci/delivery_metrics, policy/own_language, queen/priority, queen/review_valve, tri/catalog/health, tri/t27b/{dogfood,fuzz,fuzz_oracle,reduce,steward}),
gen/rust/tri/t27b/check_budget.rs, andgen/c/numeric/gf16.c+gen/verilog/numeric/gf16.v. The gf16 pair is left to #6996 item 5.The other direction (commit 2af2517a4, closes this issue).
tools/l2_regen_check.pynow also checks a tracked copy when its spec, or any spec in itsuseclosure, is added, modified or deleted in the PR. gen-c splices imports, so a changed import changes the copy. Copies of specs the PR did not touch are not charged to it, so a gen-c change does not turn every PR red.--allchecks every tracked copy. The workflow file is unchanged:--listnow prints these copies, so the step builds t27c for them.Controls, old script vs new, in a throwaway worktree:
DENYchanged in specs/policy/own_language.t27, no copy regenerated--allThe script is foreign Python: an edit to an existing tool, listed in
tools/policy/foreign-exceptions.txt, and the PR carries the owner-approved-foreign label. The debt is moving the L2 rule into a spec, asspecs/policy/own_language.t27did for the Only-t27 gate.Merge origin/master into claude/l2-rule-in-spec-land (Refs #7113)
policy: the L2 GENERATION rule lives in specs/policy/l2_generation.t2…
t27b: tail expressions and discarded calls as the reference prints th…
ci(spec-guards): published figures are red only for a defect of the c…