Skip to content

fix(ci): the seal gate decided from history CI does not have - #2445

Merged
gHashTag merged 1 commit into
masterfrom
fix/seal-needs-history
Aug 22, 2026
Merged

gHashTag merged 1 commit into
masterfrom
fix/seal-needs-history

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

Refs #2325.

The gate classifies a broken seal as dangling (spec was committed then deleted — remove the seal with it) or phantom (spec appears in no commit — find it or drop the seal), deciding from git log --all.

seal-coverage.yml used a bare actions/checkout@v4 — depth 1. A one-commit clone has no history, so the dangling arm could never fire and every seal printed as phantom: wrong class, wrong prescribed repair, on the only output the gate prints. l1-traceability.yml and now-sync-gate.yml already set fetch-depth: 0.

Measured in a real --depth=1 clone, three ways:

dangling phantom no-spec-path stale
full history 89 0 5 191
shallow, before fix 0 89 5 191
shallow, after fix 89 0 5 191

Two changes, because either alone is half the repair: fetch-depth: 0 in the workflow, and _ever_existed now detects a shallow checkout and returns the milder classification rather than answering a question it cannot see.

A gate that reads its verdict out of a truncated instrument is the broken-ruler error — and this function's own docstring records the author falling into it once already, when an earlier pass overstated fivefold.

🤖 Generated with Claude Code

check_seal_coverage.py classifies a broken seal as `dangling` -- the spec
was committed and later deleted, "remove the seal with it, or restore
both" -- or `phantom` -- the spec appears in NO commit, "find the spec or
drop the seal". It decides from `git log --all`.

seal-coverage.yml used a bare `actions/checkout@v4`, which is depth 1. A
one-commit clone has no history, so the dangling arm could never fire and
every seal printed as phantom: wrong class, wrong prescribed repair, on
the only output the gate prints. l1-traceability.yml and now-sync-gate.yml
already set fetch-depth: 0, so the omission here was load-bearing.

Measured in a real --depth=1 clone, three ways:

    full history          dangling 89, no-spec-path 5, stale 191
    shallow, before fix   PHANTOM  89, no-spec-path 5, stale 191
    shallow, after fix    dangling 89, no-spec-path 5, stale 191

Two changes, because either alone is half the repair: fetch-depth: 0 in
the workflow, and _ever_existed now detects a shallow checkout and
returns the milder classification rather than answering a question it
cannot see. A gate that reads its verdict out of a truncated instrument
is the broken-ruler error -- and this function's own docstring records
the author falling into it once already, when an earlier pass overstated
by fivefold.

Refs #2325.
@gHashTag
gHashTag enabled auto-merge (squash) August 22, 2026 17:48
@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

This notebook contains session context, decisions, and artifacts for this work.

@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-22 17:49:04 UTC

Summary

Status Count
Total Open PRs 5
PRs with Failing Checks 1
PRs with All Checks Green 4
READY 3
FAILING 1
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=e905ba8eeb53 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag
gHashTag merged commit 39876e3 into master Aug 22, 2026
22 of 24 checks passed
gHashTag added a commit that referenced this pull request Aug 22, 2026
--update-baseline writes `name | kind | detail` and reading it back kept
only the name, so a baselined entry was a permanent, kind-blind
exemption.

Measured on the real tree: 58 baselined names are no longer in the bad
set, and nothing computed that. 56 are genuine repairs the gate never
mentioned. Two seal FILES are gone outright -- FpgaEmission.json and
radix_economy.json, both admitted as `stale`, whose prescribed repair is
"re-seal it". Deleting a stale seal destroys the reproducibility record
and the gate said nothing at all.

Three classes now: CHANGED (a baselined seal in a different state, whose
repair is a different repair -- `stale` says re-seal, `dangling` says
restore or remove), DEPARTED (the file left the tree), and NOTE (repairs
to record). No file-format change: the writer already emitted the kind
and the reader was throwing it away.

MOVEMENT INSIDE {phantom, dangling} IS DELIBERATELY SILENT. That pair is
decided from git history, which a shallow checkout does not have --
_ever_existed concedes exactly that -- and 15 baselined entries sit at
`phantom` only because the ledger was written while CI ran on a depth-1
clone. #2445 gave it history, so those 15 read `dangling` now. That is
the instrument being fixed, not the tree drifting. Breaking the
suppression prints exactly those 15 rows, which is how the number was
confirmed rather than assumed.

The pre-existing "FAIL: 136" listing is NOT short-circuited -- the new
lines print above it and the verdict is unchanged, exit 1 as before.
That was the acceptance bar, because this gate is red on master for
reasons this patch does not touch: same verdict, strictly more
information, no new failures attributable to the change.

Controls in --self-check, all four pure and tree-free: a real kind change
is reported, a phantom/dangling move stays silent, a departure is named,
a repair is named. Disabling the pair suppression or the departure
accounting each takes it red.

Refs #2325.
@gHashTag
gHashTag deleted the fix/seal-needs-history branch October 26, 2026 06:11
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.

1 participant