Fix Pi context usage updates during tool loops

[ymichael]

What was wrong

Pi sampled context-window usage only on SDK agent_end, which fires after the entire agent run. A tool-heavy run contains multiple SDK turn_end events—each after an assistant response and its tool results—so bb's context meter stayed stale throughout the tool loop even though Pi's underlying context estimate was changing. See #2023.

What changed

• Sample and emit Pi context-window usage on every turn_end, while retaining the existing compaction_end update.
• Stop sampling again at agent_end; that event is still forwarded normally for completion and checkpoint handling, but the final turn_end already emitted the same context snapshot.
• Add a bridge regression test with an intermediate tool-result turn and a final response. It asserts that both usage snapshots arrive and that agent_end does not duplicate the final one.
• Bump HOST_DAEMON_PROTOCOL_VERSION from 141 to 142 because the bundled Pi bridge's daemon-to-server event cadence changed and enrolled daemons need to update.

How you verified

• pnpm exec turbo run test --filter=@bb/agent-runtime --force -- src/pi/bridge/__tests__/bridge.test.ts — 27 passed. The new regression fails before the fix because only the agent_end sample is emitted.
• pnpm exec turbo run typecheck --filter=@bb/agent-runtime --filter=@bb/host-daemon-contract --force — passed.
• pnpm exec turbo run test --filter=@bb/host-daemon-contract --force -- test/contract.test.ts — 38 passed.
• The full host-daemon-contract suite was also run locally: 51 tests passed and its unrelated fixed gzip-byte measurement test differed under local Node 26.3.1/zlib (payload-size.test.ts); the protocol contract itself passed.

Fixes #2023

AGENT GENERATED: by GPT-5