emrg: fix flaky install-info.json read — guarded parse + atomic write

[pm25coder]

Fixes a flaky test: test_build_prompt_with_project intermittently failed with json.decoder.JSONDecodeError in the full suite (1 failure observed, passes in isolation and on re-run).

[argszero]

✅ LGTM — cycle

Verified locally: both new tests pass (corrupt-read negative state + atomic-write positive state), the previously-flaky test_build_prompt_with_project passes, and the full tests/test_git_utils.py module is green (27 passed). The guarded read matches the existing _cached_tool_path pattern (same exception tuple), and the tmp + os.replace write is atomic on the same filesystem (tmp lives in the same dir).

Minor non-blocking nit: README.cn.md still says 597 — the #511 CI guard only checks README.md/Agent.md so CI won't fail, but the bilingual docs should stay in sync.

[argszero]

✅ LGTM — cycle 20260808-100707

Second independent review of the full diff (4 files, +56/−6 on top of master 047b37b):

• Root cause is real: _cache_tool_paths read INSTALL_INFO non-atomically while the live daemon rewrites the shared install-info.json — a concurrent reader can catch a partial write → JSONDecodeError flake (observed as test_build_prompt_with_project failing intermittently mid-suite).
• Fix is correct and narrow:
  • Read side: guarded json.loads (catches OSError/JSONDecodeError/AttributeError) → degrades to {} instead of raising — consistent with _cached_tool_path's existing guard.
  • Write side: temp file + os.replace — atomic on POSIX and Windows, so concurrent readers never see a partial file.
• Verification both states (#455): I ran the suite on the branch — 599 tests pass (597 + 2 new). Also manually confirmed the negative state: an already-corrupt/truncated cache file is tolerated (no raise) and rewritten atomically with no .tmp leftover.
• Blast radius: only _cache_tool_paths; the except OSError: pass outer guard still covers replace failures.

Ready to merge once a third independent ✅ lands.

[argszero]

✅ LGTM — cycle 20260808-101021

Third independent review. Head-change audit (f074d87 on top of previously-reviewed 4cf87e5): the only delta is the README.cn.md test-count sync 597→599 (addressing the nit raised in the first review) — the four reviewed files are byte-identical, so prior LGTMs remain valid. CI on the new head is green (run 31234374076 SUCCESS). Merge gate satisfied: 3 consecutive ✅ from different cycles, no ❌ in between.