QVAC-18790 infra: scope e2e bootstrap to deps required by filtered tests

[Victor-Rodzko]

🎯 What problem does this PR solve?

• SDK e2e consumers (packages/sdk/tests-qvac) call ResourceManager.downloadAllOnce on bootstrap and warm every registered model (~24 desktop / ~22 mobile) regardless of which tests will actually run.
• On Device Farm this is the dominant cost of each CI job — Android bootstrap regularly hits ~20–30 min, and one iOS run failed with ❌ No consumers connected within 1200s timeout while the device was still pre-loading models.
• A typical filtered run only needs a subset of those deps (e.g. --filter registry needs 0, --filter translation needs 8).

See QVAC-18790 for the full problem statement and CI evidence.

📝 How does it solve it?

• Companion framework change in tetherto/qvac-test-suite#70 delivers the producer's post-filter TestDefinition[] to the consumer's bootstrap callback over MQTT (register-ack.filteredTestIds).
• ResourceManager.downloadAllOnce gains an optional allowedDeps: ReadonlySet<string> — when provided, downloads + pre-loads are scoped to that subset; when omitted, the legacy "warm everything" path is preserved (used by the CI cache-warmer step and any older framework).
• New shared/collect-test-deps.ts helper extracts metadata.dependency / metadata.dependencies from the filtered tests (drops the "none" sentinel, dedupes).
• Both desktop/consumer.ts and mobile/consumer.ts bootstrap callbacks now accept filteredTests?: TestDefinition[], derive allowedDeps via collectTestDeps, and pass it through. undefined keeps today's behaviour, so older framework versions and the unfiltered CI cache-warmer continue to populate the full cache.

🧪 How was it tested?

Local

• Desktop, npx qvac-test run:local:desktop --filter transcription,parakeet → 🎯 Bootstrap dep-filter active: keeping 4 dep(s); 20 of 24 defined excluded, bootstrap 9.97 s (was minutes).
• iOS on a real iPhone, npx qvac-test run:local:ios --filter transcription,parakeet → keeping 4 dep(s); 18 of 22 defined excluded, bootstrap 20.2 s (was ~25 min). Full batch 62.86 s, 13 passed / 18 skipped (known QVAC-18460 iOS skips) / 0 failed.
• No-filter path (older framework / cache-warmer) verified by typecheck + bootstrap() called without args: allowedDeps stays undefined, manager warms everything as before.

CI

Two full QVAC Tests (sdk) runs against this PR (using the matching preview build of tetherto/qvac-test-suite#70), both fully green:

Job (mac-mini-m4-gpu)	Run 1 --filter translation (8 deps)	Run 2 --filter registry (0 deps)
Filter marker	keeping 8 dep(s); 16 of 24 defined excluded	keeping 0 dep(s); 24 of 24 defined excluded
Consumer-side Bootstrap completed in	4 098 ms	1 ms
Test results	52 / 52 pass	11 / 11 pass

Mobile, same two runs:

• Android run-producer: 22m12s → 2m31s (test set 52 → 11). 0/52 → 1/52 completed transition shows bootstrap ~18.5 min for 8 deps on S25 Ultra (down from ~30 min full-registry warm seen in run 25672716968). For the 0-dep run the device starts the first test ~10 s after register.
• iOS run-producer: 6m40s → 3m04s, no No consumers connected timeouts. 49 pass / 3 skip vs 11 / 0.
• CI runs: 25720111911 and 25722580067.

📌 Follow-ups before merge

• @tetherto/qvac-test-suite is currently pinned to 0.6.3-tmp.pr-70.runid-25717564206 (preview build of tetherto/qvac-test-suite#70). Once that PR lands and 0.6.3 is published, this dep moves to ^0.6.3.

[github-actions]

Tier-based Approval Status

**PR Tier:** TIER1

**Current Status:** ✅ APPROVED

**Requirements:**
- 1 Team Member approval ✅ (1/1)
- 1 Team Lead OR Management approval ✅ (1/1)



---
*This comment is automatically updated when reviews change.*

[Victor-Rodzko]

/review