#1202 Flake batch 28: sqlite3 probe, route census snapshot, walk progress, bash promotion e2e

Flake batch 28: test-only fixes, no product change.

• bash_db_hints real-probe test: on a loaded macOS runner, starting the real sqlite3 took longer than the probe's fixed 1.5 s budget, so the attempt counted as a timeout instead of an error. A timeout proves nothing about the error path, so the test retries with a fresh probe up to five times. Every attempt must still create no database, and an outcome that is neither an error nor a timeout fails at once.
• Route census in-flight pins test: the census reads the executor's running read requests with a try-lock and reports read_lease: null (unknown) when the scheduler holds that lock at that moment, by design. The test read it once. It now takes snapshots for up to 10 s until one isn't contended; the parked job keeps both pins in place and every assertion is unchanged.

Build server: route census and bash_db_hints tests passed three times (23 each), plus formatting and the Windows check.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

---

Summary by cubic

Fixes four flaky tests; no product change.

• The bash_db_hints real-probe test retries up to five times when the probe times out on a loaded runner, instead of failing because starting sqlite3 exceeded the 1.5 s budget.

• The route census test retries the snapshot until the read-lease count is uncontended, instead of reading once and occasionally seeing read_lease: null.

• The folder-walk progress test now grants the walk a 3 s budget over 80 entries, so it fits well over five entries on loaded runners and still hits the time limit on fast ones.

• The Pi and OpenCode foreground-promotion tests pass AFT_TEST_FOREGROUND_WAIT_MS through the bridge's childEnv and keep the command alive 5 s so terminal completion doesn't beat promotion.

• Adds a search descriptor for the flake batch.

• Folder-walk progress test: on a loaded macOS runner the walk fit only 5 entries in its budget. It now gets room for well over five entries, still with a bounded deadline.

• Pi and OpenCode foreground-promotion e2e tests: the engine reads AFT_TEST_FOREGROUND_WAIT_MS from its own environment at spawn, so the tests now pass it through the bridge's childEnv instead of the ambient process environment. Pi's 0.2 s command could finish before the engine polled for promotion on a loaded runner (a finished result wins over promotion), so Pi now uses OpenCode's 5 s command, and both assert the server text names the 0 s window. Pi and OpenCode bash e2e files passed five times each on the build server.