fix(core): a same-directory skill name collision lands in the refreeze receipt

[Max17190]

Why

Round-5 ticket T3 (.scratch/selfext-round5/findings.md). skills_by_name keys on the frontmatter name, last-sorted wins, so two files in one skills directory declaring the same name collapse silently. --check catches it, but the same-turn refreeze receipt after the write that caused the collision said nothing, so the author reads the unchanged index line and assumes the older file is what loaded.

Stacked on #248 (the capture-snapshot fields this adds sit beside the ones #248 introduced; merging this first would conflict).

Summary

• The capture records same-DIRECTORY collisions as (name, ignored path, indexed path) on the snapshot and registry (empty for a manifest-restored registry). Cross-tier shadowing is deliberate precedence (project over global) and stays unrecorded.
• refreeze_receipt_text gains the collision clause: both declaring files, which one is indexed, and the repair (rename one; verify with openmax --check).

Test Plan

• a_same_tier_skill_namesake_is_recorded_and_precedence_is_not: two project subdirs declaring one name yield exactly one recorded collision with the sorted-last winner; the same name split across tiers records nothing. Red against a no-record variant.
• refreeze_receipt_names_a_skill_shadowed_by_a_same_tier_namesake (the ticket's red-test name): the receipt names the collision, both project-relative paths, the winner, and the verify command. Red against a clause-less receipt.
• Full workspace green (exit 0), clippy zero warnings.

Greptile Summary

Skill-collision receipts now reflect the final active definition and its actual prompt visibility. Same-tier collisions are coalesced around the final winner, project-level overrides remove irrelevant global collision history, and winners excluded by either the skill cap or prompt byte budget are reported as not indexed.

Focused Rust tests exercised the 50-skill cap, three-file same-name collision, project override, prompt byte-budget exclusion, and receipt wording. All passed, disproving the previously reported collision receipt failures.

Confidence Score: 5/5

No blocking failure remains.

The exercised collision and receipt paths produce the expected active-skill and prompt-visibility outcomes.

T-Rex Logs

What T-Rex did

• Reviewed the registry's coalescing of same-tier records and the removal of a losing-tier collision when project precedence wins.
• Verified that winner_indexed is derived from actual prompt-index inclusion, covering both the skill cap and the byte budget.
• Confirmed that agent.rs emits only when the settled flag is true; otherwise it explicitly says no file of that name is indexed.
• Observed that the executed test run produced five passing focused tests with the command, working directory, and exit code for each.

_{Ran code and verified through T-Rex}

_{Reviews (6): Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." | Re-trigger Greptile}