Separate declaration key coverage from lifecycle startup

[ymichael]

What was wrong

PR #2084 removed avoidable TypeScript fixture transforms from this test, but the unified provider-harness work in #2325 later moved it back onto the TypeScript scripted-echo bridge while preserving three serial real-process launches. The packages shard makes each tsx-backed startup compete with many other workers; the redundant third launch pushed the test over its unchanged 15-second cap. Under controlled contention, the request log reached the third process's initialize request but timed out before its thread/start request, confirming cumulative process startup as the cause rather than a polling deadline. No bridge process remained after the timed-out worker exited.

What changed

Successive capability-declaration fingerprints at one artifact hash now have deterministic coverage in bridge-launch-process-key.test.ts. The end-to-end lifecycle test retains the minimum real-process boundary: two equal declarations share one initialized bridge, then one changed declaration initializes a distinct second bridge. This keeps assertions for hangs, process reuse, and process distinction while removing the third serial launch. The 15-second timeout is unchanged. This is test-only; there is no production, wire, CLI, guide, or protocol-version change.

I rejected a timeout increase because scheduler contention is only the reproduction condition. I also kept two real bridges instead of replacing the lifecycle boundary with an injected process factory, because two processes are the minimum needed to prove end-to-end distinction.

How you verified

• Unchanged baseline, unloaded: focused test passed in 596ms.
• Unchanged baseline under 192 and 384 CPU competitors: passed in 7.484s and 11.895s.
• Unchanged baseline under 576 CPU competitors: reproduced the exact failure at 15.009s with Test timed out in 15000ms.
• Changed test under the identical 576-competitor condition: passed in 11.052s with the same timeout; no scripted bridge process remained afterward.
• Changed test, unloaded: 330ms; 10/10 fresh Turbo/Vitest repetitions passed (about 0.31s when reported).
• pnpm exec turbo run test --filter=@bb/agent-runtime --force --output-logs=full -- src/bridge-launch-process-key.test.ts (2 passed).
• pnpm exec turbo run test --filter=@bb/agent-runtime --force --output-logs=full -- --project '@bb/agent-runtime:isolated' src/runtime.process-lifecycle.test.ts (35 passed).
• pnpm exec turbo run test --filter=@bb/agent-runtime --force --output-logs=full (21 files, 317 tests passed).
• pnpm exec turbo run typecheck --filter=@bb/agent-runtime --force --output-logs=full (passed).
• pnpm exec turbo run build --filter=@bb/agent-runtime --force --output-logs=full (upstream generation passed; this package has no build task).
• pnpm exec turbo run lint --filter=@bb/agent-runtime --force --output-logs=full (this package has no lint task), plus targeted oxlint, oxfmt --check, and git diff --check (passed).

AGENT GENERATED: by GPT-5.6-Sol