feat(release): 2 of 10 bundled artifacts were signed but undescribed for 58 releases

[avrabe]

Verify-Filter: (has-tag "manifest")

Filter narrowed from (or (has-tag "manifest") (has-tag "release")). The
broader form selected FV-FALCON-RELEASE-001, whose three gh / cosign
steps exit rc=127 on the gate runner and had never been executed — that
artifact carries no falcon tag, so the default sweep never matched it. That
is a real latent defect, not something to filter around, and it is fixed in
#374 by classifying release-verification commands bench-only by shape
(4 steps across 2 artifacts). This PR narrows to the artifact it actually
introduces; once #374 lands the broader filter would pass too.

jess asked on #202 which cascade artifact to fuse. The bundle could not answer.

Measured on the published artifact

falcon-components-v1.135.tar.gz, pulled from the falcon-v1.135.0 release:

	count
entries in the cosign-signed SHA256SUMS	10
components declared in manifest.json	8
signed but undescribed	2

falcon-cascade-stream-composed-v1.135.wasm
falcon-cascade-stream-fused-v1.135.wasm

Why that's an ambiguity and not untidiness: one directory holds
falcon-cascade-v1.135.wasm — the cargo-component cascade that gets fused and
lowered — beside falcon-cascade-stream-{composed,fused}-v1.135.wasm, which are
P3 async stream variants: a different artefact class with a near-identical
name. Asked "which of these did you verify, and which do I fuse?", a consumer
had only the filename, because the machine-readable manifest listed neither.

The attestation was never broken

Stated explicitly because which failure this is matters. Inside the unpacked bundle:

$ shasum -a 256 -c SHA256SUMS
./falcon-attitude-v1.135.wasm: OK
... all 10 ...
exit 0

So SWREQ-RELAY-STREAM-P15 is genuinely satisfied — the stream artifacts are
bundled and are covered by the signature. What drifted, silently, across 58
releases is the bundle's self-description, because nothing compared the two files.

Fix: describe, don't remove

manifest.json now carries kind:

{ "name": "falcon-cascade", "file": "falcon-cascade-v1.136.wasm",
  "kind": "flight-component",   "sha256": "...", "bytes": ... }
{ "name": "falcon-cascade-stream-fused", "file": "falcon-cascade-stream-fused-v1.136.wasm",
  "kind": "p3-stream-artifact", "sha256": "...", "bytes": ... }

Telling them apart becomes a field lookup instead of a guess from a filename.

Removal was the originally agreed fix and is rejected on evidence: it would
contradict SWREQ-RELAY-STREAM-P15 (which requires precisely that they be
bundled and attested), falsify FV-RELAY-STREAM-015, and would not answer the
question actually asked — which is "what IS this file", not "why is it here".

Drift guard, tested both ways

negative (the exact v1.135 shape, reproduced deliberately):
  ERROR: signed by SHA256SUMS but undescribed in manifest.json:
         falcon-cascade-stream-fused-v1.136.wasm            exit 1
positive:
  manifest/SHA256SUMS agree on every bundled artifact       exit 0

A guard only ever seen to pass is indistinguishable from one that cannot fail, so
the negative case is the bug that motivated it, not a synthetic stand-in.

A correction worth recording

This was first mis-diagnosed as something far worse — "9 of 10 signed files are
missing from the release". Listing release assets through a name filter of
wasm|SHA256|manifest excludes falcon-components-v*.tar.gz, whose name contains
none of those; checking each SHA256SUMS entry against that already-filtered list
then reported 9 MISSING. The tarball was there the whole time — the filter was
applied twice, and the second use inherited the first one's blind spot.

The real defect is smaller, different in kind, and was only visible after
unpacking the actual published artifact rather than reading the script that
produces it.

Evidence

rivet validate                                       exit 0
FV-FALCON-REL-002: 4/4 steps PASS
shasum -c SHA256SUMS (published v1.135 bundle)       10/10 OK, exit 0
guard negative / positive                            exit 1 / exit 0

Two-commit rule: this PR carries code, so nothing goes past implemented.

Refs #202

🤖 Generated with Claude Code

https://claude.ai/code/session_01HvusAXYbHLyv3uTzfBcMbG