fix(server): make debug-path startup receipt durable so logger.separation tests never time out on it

[danshapiro]

Problem

test/integration/server/logger.separation.test.ts flaked on the cloud backend with a 30s timeout waiting for the Resolved debug log path marker line. Root cause: the marker was routed through pino's multistream into rotating-file-stream, which opens the file lazily and buffers writes — a short-lived probe process exiting before the async open completed silently discarded the line, so no file-content wait could ever see it. (The earlier 5s→30s wait widening on main only changed how long we wait for a line that was never written.)

Verified empirically: the immediate-exit probe (import logger then process.exit(0)) loses the marker 20/20 at the base commit.

Fix

In server/logger.ts, the one-time resolved-path receipt is now appended synchronously (fs.appendFileSync) during createLogger(), gated on isLevelEnabled('info') to preserve the exact prior suppression semantics, with an identical JSON shape (byte-compatible with what the tests parse; no pid/hostname). The rotating stream continues to handle the ongoing verbose flow unchanged. A narrow try/catch degrades a marker write failure to a diagnostic warning instead of crashing startup.

Tests

• 2 new unit tests pin the receipt's exact shape (incl. env/version presence/omission) and LOG_LEVEL=warn suppression.
• 1 new integration test proves a process that exits immediately after import still leaves the receipt (and exits 0).
• The separation tests (concurrent/dist-source/explicit-instance) now also assert a stream-routed, per-instance proof line in each file — with cross-contamination negatives re-read after both positive waits — so they still prove each process writes through its own rotating stream into its own file (independent review caught that the out-of-band marker alone had silently removed that coverage).

Verification

• Full suite on the cloud backend (the flake's venue) green at HEAD: all shards passed, target file 6/6 inside the full-suite cloud run; electron 350/350 locally.
• Development ran through the complete the-usual workflow: load-bearing validation (4/4 verified, incl. empirical 20/20 marker-loss at base), independent plan review (2 rounds), task + whole-branch reviews (APPROVED), independent delta review (2 rounds, final PASSED) with a focused repair episode (2 rounds PASSED).

Fixes the flake without touching any wait durations.