Test efficiency phase 2: measured-cost tiers and suite dedup

[backryun]

Follow-up of #12539 (phase 1 landed as #12588: all 4,263 flat tests/unit files relocated into per-prefix subdirs, flat = 0).

Phase 1 findings (measured in worktree)

• Exact-duplicate files: 0.
• Shared test names across files: 190 (mostly generic names; a few real overlap candidates, e.g. antigravity usage-fetcher vs usage-service sharing 6+ quota-parsing cases).
• Cost-signal scan (spawn/sleep/db/net/timing keywords) per top dir: db-touching is pervasive (combo 83, db 102, misc 55 files); spawn concentrates in cli/build/db; only 74 files in 12 dirs are fully signal-free — too small for a meaningful 'quick' tier.
• Full test:unit single pass exceeds local time budget (killed after 2h at ~25% progress); CI unit gate hits the 4800s ceiling (🔴 Release branch not green: release/v3.8.51 #12581).

Proposed scope

1. Measure: per-directory wall times (CI reports or sampled local runs) to rank the slowest dirs.
2. Tier: add an opt-in fast tier (e.g. test:unit:quick = all dirs minus measured-heavy set). Additive only — do not change existing test:unit/CI wiring.
3. Dedup: merge verified-overlapping suites case by case (starting with the antigravity usage pair), deleting only with passing-suite proof.

Non-goals

• CI shard/runner redesign (belongs to CI/infra, cf. 🔴 Release branch not green: release/v3.8.51 #12581).
• Changing default test:unit coverage.

[backryun]

Measurement done (worktree, concurrency 10):

dir	files	tests	wall
issue	89	338	167s
combo	164	1330	158s
compression	242	1574	150s
provider	194	1428	141s
misc	220	1804	98s
db	128	1172	96s
model	94	673	84s
cli	191	1229	79s
quota	102	904	46s
ui	275	209	5s

Slow tier = the 6 dirs at >=95s (combo, compression, provider, misc, db, issue).
Implemented: test:unit:quick (branch chore/test-tiers-12589, stacked on #12588) — additive only, CI untouched. Remaining failures in measured runs are all pre-existing env/timing flakes (binaryManager /tmp symlink, runner-janitor env, audit-timeline tz, 9147 event-loop under load).

[backryun]

Phase 3 (dedup) finding: NO merges justified.

• Exact-duplicate files: 0.
• Top overlap candidates examined file-by-file (antigravity usage-fetcher vs usage-service; build vs check cognitive-complexity; build vs check type-coverage; cli vs codex windows-spawn): all are complementary suites pinning different units/behaviors (fetcher.ts vs usage.ts; parse vs evaluate; spawn vs quoting angles). Shared test names reflect parallel contracts, not copies. Merging any of them would delete coverage.
• Conclusion: the suite's size comes from breadth (355 providers x(executors, translators, OAuth, catalogs)), not duplication. Further count reduction must come from retiring whole provider surfaces, not test merging — out of scope here. Closing phase 3 with no deletions.