ci: add opt-in paired race cache seed measurement

[XuPeng-SH]

Problem

UT cache seeding is default-off after #455. The Shanghai amd64-mo-shanghai-8c16g runner has no Docker daemon, and there is no controlled measurement of whether cache acquisition/import pays for itself over the complete race suite.

Change

Add a manual cold/warm A/B canary on the actual Shanghai runner class with immutable MatrixOne main SHA, CI checkout and builder digest. A disables seeding; B uses an opt-in daemonless registry transport that imports only the trusted producer's final cache COPY layers. Normal callers keep the existing Docker transport and seed remains default-off.

Each pair uses matched initial cache state, alternating AB/BA order, and records seed acquisition/import/cleanup plus clean/config/full race wall time, cgroup CPU/memory, disk and exact test/package outcomes. The comparison rejects identity drift, missing tests, partial imports and incomplete cleanup. Coverage is excluded.

Validation

• Linux Python 3.12: 43 importer/cleanup/registry tests and 11 canary tests pass, including strict layer admission, compressed/uncompressed digest verification, malformed DEFLATE, cancellation, populated-cache preservation, forced EXDEV warm restore and exact outcome comparison.
• actionlint 1.7.12 passes the reusable workflows, validation workflow and MatrixOne caller; whitespace checks pass.
• Live task-owned Shanghai 8 CPU / 16 GiB probe, using producer 3ac87c30625fc082391e5384a4c94e50a2916cf6 and pinned image sha256:8137d222d3a3b29d119173639cea98931168ab7d886bf882894391f7804d5d88: state=seeded, producer ok, modules seeded, cleanup complete; 64,500 files / 15,819,961,352 bytes imported in 1,029.392s. A forced go test -race -count=1 smoke executed its body and passed.
• No OOM/OOM kill, but memory reached about 16.4GB with 64 memory.max events. The diagnostic pod was deleted after collection.
• Independent GPT-6 medium design and implementation review passed; all findings were addressed.

Limits

The live result proves capability, not speedup. Seeding consumed 1,029s of the 1,080s work budget and has little memory headroom. Full cold/warm MatrixOne race A/B has not run and this PR does not enable rollout. A thin MatrixOne manual entrypoint, pinned to the reviewed CI implementation commit, is required after this PR is merged so it can inherit the existing UT credentials.

See docs/race-seed-canary.md and docs/race-seed-validation.md for the measurement contract and evidence.

[XuPeng-SH]

Reviewed exact head 23eb3cc728e27db9d3e680bf78f69a44911a0bf5; self-authored PR, so COMMENT only.

[P2] Cold B admission must reject skipped module import.

scripts/race_seed_canary.py:321-326 validates overall seed status, producer health, cleanup, and imported build-file count, but ignores module_state. The importer can return state="seeded" with module_state="insufficient-space" after build-cache import succeeds but module import is skipped (actions/seed-go-caches/seed.py:489-515). B can therefore omit module acquisition/import while still being accepted as a valid comparison, making measured treatment and acquisition cost depend on available disk.

Require module_state="seeded" for cold B, allow only the expected preserved-populated state for warm B, and add regression coverage.

The 43 importer tests, 11 canary tests, and actionlint passed. Full Shanghai cold/warm A/B runtime evidence remains outstanding.

[XuPeng-SH]

Re-reviewed exact head 20a53a824c4526942ca5e570fb4f0c865f8c6492; self-authored PR, so COMMENT only.

PASS — the previous P2 is fixed at both sample admission and comparison: cold B now requires module_state=seeded, warm B requires preserved-populated, and regression coverage rejects missing, insufficient-space, and mismatched states. No additional actionable P0–P2 findings across the complete PR.

CI passed on an identical tree, including actionlint and 55 relevant tests. Evidence gaps remain for the full cold/warm race A/B and separately delivered MatrixOne caller integration; retain default-off rollout and do not infer a speedup or rollout conclusion yet.