verify(v1.133): record the release evidence — and why NOTHING is promoted yet

[avrabe]

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

Scoped: the full 156-artifact sweep is now killed at the job's own
timeout-minutes: 90 (this PR died at 1h30m23s, #344 at 1h30m21s). This PR
only edits artifact descriptions — no steps change — and (has-tag "oci")
matches 4 artifacts including all three touched here, so the scoped sweep
exercises exactly what changed. Timeout raise tracked in #350 / fixed in #342.

falcon-v1.133.0 shipped and every falsification criterion in the tag passed. This records that evidence against the three verification artifacts.

It promotes nothing. Each requirement’s own verification-criteria are broader than "the release worked", and promoting on the evidence I happen to have rather than the evidence the requirement asked for is exactly the failure this process exists to prevent.

What the release proved

Measured on the published artifacts, not the workflow’s word:

check	result
header 0061736d0d000100 (component, not core module)	6/6
memory.grow	0
wasi:* imports	0
embedded WIT namespace pulseengine:falcon-cascade	6/6
legacy falcon:cascade	0
OCI annotations	9 each, real descriptions

cosign verify → subject release.yml@refs/tags/falcon-v1.133.0, githubWorkflowSha 7809e93d — the exact commit the pre-tag gate was green on. That closes the loop between "CI was green" and "this is what people pull".

REL-P01 observed live: 18:03 draft → 18:38 draft with 12 assets attached → 18:53 published with 20.

The 18:38 sample is the load-bearing one: twelve assets attached and still invisible to consumers. That is precisely the v1.131.0 window, now unreachable rather than merely unlikely.

Why each artifact stays implemented

OCI-P02 — criteria carry a kill-criterion owned by gale/meld: with every stage converted, meld fuse --memory shared --address-rebase must accept the fused cascade. Not run. Zero memory.grow is necessary for that fuse, not sufficient — it was the known blocker (gale#89, meld#299), not a proof. Promoting on bytes alone would assert the very outcome the kill-criterion exists to test.

OCI-P04 — (a) published + anonymously pullable → met; (b) wasm.directory registered with [[interface]] entries → not met; (c) a component page’s export links resolve into an owned namespace, "verified by fetching the live pages, not by inspecting the workflow" → not met. (b) and (c) need the registry/pulseengine.toml follow-up for the nested paths.

REL-P01 — criteria say verification is by "DELIBERATE FAILURE INJECTION, not by reasoning": (a) no published release after a broken component build, (b) the guard still blocks, (c) an unbroken run publishes correctly. This release discharges (c) only.

The distinction this PR is making

The artifacts are verified. The requirements are not. Those are different claims, and this release only supports the first.

Evidence

• rivet validate PASS, exit 0 — v0.19.0, the version the gate pins
• gate on these artifacts: 7 steps PASS, 0 FAIL
• no status changes; no code

🤖 Generated with Claude Code

https://claude.ai/code/session_01HvusAXYbHLyv3uTzfBcMbG