fix(targethasher): collapse apparent-name bzlmod external files via repo mapping

[yushan8]

Summary

Why?

bazel query prints repos visible to the main module by apparent name (@name//...), but bzlmod marker hashes are keyed by canonical name and the collapse only recognized @@canonical//... labels. Files in such repos missed both the git known-hash map and the marker collapse, so the hasher read them from disk. On the java repo that was 26k files and about 40 GB (large jars, zips and container images), making targethasher.FromProto the largest part of nativeGraphRunner.Compute.

What?

• Add bazel.RepoMapping, which runs bazel mod dump_repo_mapping "" to get the apparent-to-canonical repo map, and HashConfig.RepoMapping to carry it.
• Resolve @name//... labels through the mapping in HashExternalTargetsBzlmod, so they collapse to the same marker-based hash as the equivalent @@canonical//... label. A mapped repo with no marker is left alone instead of erroring; fullHashRepos matches either name.
• nativeGraphRunner.Compute fetches the mapping for bzlmod repos and emits repo_mapping_duration. A failure fails the compute, like a marker read failure, so hashes stay deterministic.
• query-bench mirrors Compute for Bzlmod repos (--bzlmod, on by default) and logs the duration of each phase through zap: query, git file hashes, marker read, repo mapping and hasher. It also gains --cpuprofile for the hasher phase, and the graph JSON dump is opt-in via --dump.

Compatibility: collapsed external files now hash from their marker instead of their content, so their hashes (and the graph hashes above them) change. Cached graphs computed before this change will report those external targets as changed once. Cache keys do not encode an algorithm version, so existing entries need to be invalidated or the one-time diff accepted.

Issue

None.

🤖 Generated with Claude Code

Test Plan

Measured with example/cmd/query-bench on a 1.3M-target bzlmod monorepo (about 640k files, 18.8k repos) with a warm Bazel server. The target count is unchanged at 1,321,910.

• targethasher.FromProto: 65-68s before, 10.6-12.7s after.
• Total Compute: about 2m05s before, about 1m05s after.
• The repo mapping adds about 3.5s.

Issues

Stack

1. @ fix(targethasher): collapse apparent-name bzlmod external files via repo mapping #322
2. fix(itg): wire bzlmod external-target collapse into incremental graph updates #324