[Nic-dorman]

Linear issue

Closes V2-1828

Summary

build.rs declared cargo:rerun-if-changed for three hard-coded paths: .git/HEAD, .git/refs/heads and .git/packed-refs. Cargo treats a rerun-if-changed path that doesn't exist as changed on every build. So whenever one is missing, every cargo invocation reruns the build script and recompiles ant-node from scratch.

Where it bites:

• Git worktrees. There .git is a file, not a directory. Locally, CARGO_LOG=cargo::core::compiler::fingerprint=info cargo check reports stale: missing ".git/HEAD" on every run.
• CI checkouts. One of the paths is missing there too, probably packed-refs in the shallow clone.
  • On ci: consolidate PR CI, run every test target, build the shipped binary #269's run for a8e4cb1, a prebuild step had already built all 10 integration test binaries. Each cargo test --test X afterwards still printed Compiling ant-node and spent about 2m20s rebuilding the identical binary (same hash). Its tests then ran in under a second.
  • Today's ci.yml pays this on every cargo test --test X step, which is why each PoC step takes 2–5 minutes (ubuntu ~2, macOS ~3, Windows ~4.5) for sub-second tests.

Fix: ask git for the real directories (git rev-parse --git-dir and --git-common-dir), and emit rerun-if-changed only for paths that exist. ANT_GIT_COMMIT is still refreshed when HEAD or a branch ref moves.

After review (8acb6c7): with every ref packed and no refs/heads directory, the first version watched nothing for branch refs. The next commit creates refs/heads/<branch> without touching HEAD or packed-refs, so ANT_GIT_COMMIT could go stale.

• The watch-set logic now lives in build_support/git_watch.rs, shared by build.rs and a new test target through #[path].
• When refs/heads is missing, it watches the nearest existing ancestor inside the common dir (normally refs/). Cargo scans that recursively, so it sees the loose ref appear. Every emitted path still exists.

Split out of the CI consolidation (#269, V2-1811) so it can land on its own. #269 carries the same commit and will drop it on rebase once this merges.

Risk tier

• T0 — docs / tooling / CI / pure UX-output. Repo CI only.
• T1 — client-only, no network-facing behavior change. CI + prod compat smoke.
• T2 — node/client logic with behavioral surface, no protocol/format/economics change. Dev testnet + ADR.
• T3 — protocol / storage format / payments / routing. T2 evidence + adversarial testing.

Compatibility

• Wire: none
• Storage: none
• API: none. The only output of build.rs is the ANT_GIT_COMMIT env var, which is unchanged.

Semver impact

• breaking
• feature
• fix

Test evidence

• Before (main, in a worktree): every cargo check reruns the build script and rechecks ant-node, with the fingerprint log reporting stale: missing ".git/HEAD".
• After: the first cargo check reruns the script, and the second is fresh (Finished … in 0.92s, nothing marked dirty). The emitted paths are the real ones: the worktree's HEAD, plus the common dir's refs/heads and packed-refs.
• A new commit still updates ANT_GIT_COMMIT: after committing, the next cargo check reran the script and emitted the new short hash (e0afed1).
• New tests/build_script_git_watch.rs (3 tests, real git repositories in a temp dir), run by a new ci.yml step:
  • a branch commit changes a watched path
  • refs packed and refs/heads removed: the next commit leaves HEAD and packed-refs untouched but still changes a watched path
  • in a worktree, every watched path exists and a commit is seen
• All 3 pass locally (Windows). With the first version's logic put back, the packed-refs test fails with "the commit changed no watched path, so ANT_GIT_COMMIT would go stale", so it catches the bug review found.
• After the change, a second cargo check is still fresh (nothing marked stale or dirty).
• cargo fmt --check is clean, and clippy raises nothing in build.rs.
• This PR's own CI run is the live check. It uses today's per-target ci.yml, so the Run v12 … / bootstrap-stall / crash-safety steps should drop from 2–5 min each to well under a minute.

New dependency

none

ADR

n/a

Mitigation / rollback

Revert the commit. The worst case of a bug here is a stale ANT_GIT_COMMIT in a local build; release builds come from a fresh checkout.

🤖 Generated with Claude Code

[MaidSafe-Hermes]

REQUEST_CHANGES — at c04cb27. Coverage: 2/2 seats reviewed (sol-1 gpt-6.1-sol xhigh; glm-1 z-ai/glm-5.3); they disagreed. I confirmed sol-1’s blocker with a Git reproduction: when refs are packed and refs/heads is absent, the build registers no watch for that directory; a later commit creates refs/heads/ but leaves HEAD and packed-refs unchanged, so Cargo can retain stale ANT_GIT_COMMIT. Please watch an existing ancestor when refs/heads is absent (without emitting a nonexistent path) and add a regression test for this transition. Location: build.rs:44. This is separate from the pending CI and Security Audit result.

[MaidSafe-Hermes]

APPROVE — reviewed head 8acb6c7. Coverage: 2/2 seats reviewed (sol-1 gpt-6.1-sol xhigh; glm-1 z-ai/glm-5.3); both seats approve. The prior stale ANT_GIT_COMMIT case is fixed by watching the nearest existing ancestor when refs/heads is absent, and the new packed-refs regression test covers it. Build, test, format, clippy, MSRV and other CI checks have passed. Security Audit remains failed; it is separate from this code-review verdict and remains relevant to merge readiness.