Report first-party skill names on skill_used, bucket the rest as custom

[TheGreatAxios]

Summary

• skill_used carries skill_name: a first-party corbits-skills name from a closed 16-name allowlist, or custom for unknown, project-local, and plugin-authored skills
• TELEMETRY.md documents the allowlist and drops the no-name claim; the event allowlist carries skill_name and nothing else identifying

Verification

• bun run typecheck, bun run build, and bun run check pass (7134 pass, 0 fail)
• Classification unit tests pin the 16-name set and cover the employer-named-skill leak case via the serialized wire payload

Fixes CL-7326

[linear-code]

CL-7326

[TheGreatAxios]

Audit: the data structure is fine — closed allowlist + custom bucket, skill_name on the event, captureSkillUsed at the use_skill boundary. The rationale is wrong and must be rewritten before merge:

The classify.ts comment and TELEMETRY.md both claim the allowlist covers "user-invocable" skills and that bundled skills marked user-invocable: false (native-integration, native-runtime, ponytail, idiot-proof — 4 names) are "deliberately excluded". Ground truth on main: 11 skills carry user-invocable: false, and 7 of the names in this allowlist (git-rebase, git-worktrees, linear-issue-workflow, opsh, philosophy, style, typescript) carry that flag. So the comment describes a set that does not match the code. Correct framing: first-party corbits-skills names are reportable by name because they are our own strings (nothing user-authored leaks), and everything else buckets to custom. Please also document why slash-opt-out is not telemetry-opt-out: user-invocable: false means "no slash command", not "must not be named in telemetry" — the privacy property comes from first-party authorship, not from invocability.

[TheGreatAxios]

Audit (fleet builder): rationale rewrite — data structure stands, words were wrong.

Ground truth confirmed: 11 bundled skills carry user-invocable: false. Seven of them (git-rebase, git-worktrees, linear-issue-workflow, opsh, philosophy, style, typescript) are real use_skill recipes that merely opt out of slash synthesis (PRODUCT.md: they stay use_skill only), so they belong on the telemetry allowlist — slash-opt-out is not telemetry-opt-out. Only the four bake-only background skills (idiot-proof, native-integration, native-runtime, ponytail) are excluded, because they are baked into agent prompts rather than invoked as skills. Fix pushed: classify.ts comment + TELEMETRY.md now say first-party-skills-reportable-by-name with the 7/4 reconciliation spelled out (was: wrong user-invocable:false-exclusion story covering only 4). Test pins all four bake-only names to custom. Allowlist contents unchanged.