fix(session): report daemon health probe failures

[samuela]

Problem

hunk session currently collapses every /health failure to null. If the port still accepts TCP connections, the CLI then reports that another process owns it—even when a legitimate Hunk daemon is merely slow or returned a diagnosable HTTP/payload error.

This discards the evidence needed to distinguish a timeout, HTTP failure, malformed response, or transport error.

Change

• add an internal typed health-probe result that retains failure kind and elapsed time
• preserve the existing nullable readSessionBrokerHealth and boolean isSessionBrokerHealthy contracts
• include the terminal probe diagnostic only when a reachable listener ultimately causes hunk session to fail
• bound and sanitize runtime-generated transport messages, without reflecting response bodies
• keep successful probes and expected no-daemon session list calls silent

The retry policy and 500 ms timeout are intentionally unchanged; this PR adds observability rather than changing daemon lifecycle behavior.

Validation

• bun test src/session/broker/brokerLauncher.test.ts src/session/agent/commands.daemon.test.ts — 16 passed
• HUNK_TEST_SHARDS=1 bun run test — 3508 passed, 12 skipped, 0 failed
• bun run typecheck
• bun run lint
• bun run format:check
• bun run deps:check
• bun run changeset:status
• manual source CLI check against a loopback listener returning HTTP 503; verified one bounded diagnostic on stderr and no stdout

Tested on NixOS/Linux with Bun 1.3.13. The delayed-response timeout test is skipped on Windows because of the existing Bun loopback-abort limitation documented by the session probe tests.

[greptile-apps]

PR author is not in the allowed authors list.

[vercel]

@samuela is attempting to deploy a commit to the Modem Team on Vercel.

A member of the Team first needs to authorize it.

[samuela]

i had a peculiar instance today where /health took >500ms, resulting in a timeout when using the skill. proposing adding some logging here to increase observability