conformance-ct: environment hygiene — sibling pin drift, cold-bootstrap margin, lock hash churn

[lann]

Bundle, one area: the sibling-checkout arrangement's operational edges (all interim until component-test publishes to registries).

1. "At or past the pin" is unenforced in either direction. Nothing local compares the sibling's HEAD to .component-test-rev (the pin's only consumers are the five CI clone steps); _ct-tools keys bin/ reuse on the sibling's HEAD, whatever it is. Newer-than-pin: local green does not imply CI green (path deps leave no Cargo.lock trace unless the dep graph changes). Older-than-pin: loud but cryptic (e.g. a pre---suite-artifact ct-runner rejects the flag run-composed passes) with no message naming the pin. The pin moved four times on its first day — the coupling is hot and the latitude will be exercised. Fix shape: a drift check in _ct-tools (warn on mismatch, name the rev and the remedy).

2. Cold-bootstrap margin vs the timeout/cache-save interaction. A timeout kill skips cache-save, so an undersized budget never converges (bacdf58's own analysis). Current numbers: 30-minute job timeout, measured full-cold 16m31s, pin-bump cold-ish 11m51s — ~13min margin, eroding with corpus growth. Secondary: the rust-cache restore is conditioned on the artifact-cache hit, so an artifact hit + tools-cache eviction rebuilds the wasmtime-embedding runner with no registry cache (slow-only). Worth a margin note and periodic re-measurement.

3. artifact_sha256 churn in lock diffs. Upstream ruled artifact hashes are provenance, not identity (component-test#44): lock --check ignores them, but lock-update still rewrites the field with a machine-dependent value (rustc embeds absolute paths; the SDK is an out-of-workspace path dep), so every lock regeneration on any machine carries a meaningless hash-line diff — diluting "the diff is the review surface". Options: stop committing the field, normalize it, or teach lock-update to preserve the committed value when the inventory is unchanged.

[lann]

PR #320 resolves item 1 structurally: component-test is now a git dependency pinned by rev in [workspace.dependencies] (tools cargo-installed from the same pin, rev read from Cargo.lock), so the "at or past the pin" latitude no longer exists — local builds use exactly the pin or an explicit [patch]. Items 2 (cold-bootstrap margin) and 3 (lock artifact_sha256 churn) remain.

[lann]

Correction to the note above: #320 was a duplicate of the maintainer's earlier #319 and is closed — #319 is the PR that resolves item 1 structurally (same mechanism: git dependency pinned in [workspace.dependencies], tools cargo-installed at the Cargo.lock rev). Items 2 and 3 remain.

[lann]

Re-verified against main post-#319: item 1's sibling-checkout premise is gone — component-test is a git dependency pinned in Cargo.toml/Cargo.lock, _ct-tools installs at exactly the locked rev, and the recipe now cross-checks the npm pin (jco/package-lock.json) against the cargo pin, failing on skew. That covers the drift class item 1 described, in the new arrangement's terms.

Items 2 (cold-bootstrap margin vs the timeout/cache-save interaction) and 3 (artifact_sha256 churn in lock diffs — visible again in #324/#325's near-identical lockfiles differing only in the hash line) remain open. Item 3's cleanest fix is upstream (teach lock-update to preserve the committed hash when the inventory is unchanged, or stop emitting the field); repo-side normalization would re-implement lockfile parsing in a justfile.