perf(core): avoid duplicate first-turn filesystem discovery

[apanasenko-oai]

Summary

Remove avoidable remote filesystem work from first-turn startup.

• Reuse the project root discovered for AGENTS loading during the immediately adjacent skills warmup.
• Require an exact discovery-cwd match and never retain the hint across turns.
• Replace the serial skills marker walk with bounded concurrent discovery.
• Defer turn-diff display-root discovery until the first exact patch delta.

This is the startup-filesystem slice split out of #30632.

Measured effect

The baseline turn spent 438.6 ms in turn_diff_display_roots before the first model request.
Post-change no-tool traces contain no display-root span, proving that the entire operation is absent unless a patch diff actually needs it.

Validation

• CI is green at the current head.
• just test -p codex-core-skills repo_skill_roots passed 2 tests.
• just test -p codex-core display_roots_are_initialized_only_when_supplied passed.
• just fmt passed, with an unrelated justfile rewrite discarded.
• just bazel-lock-update completed without changing MODULE.bazel.lock.