fix(test-runtime): isolate run-graph Codex logs per run to stop flaky live-run detection

[ElonSG]

Root cause

fkst.test.run_graph invoked every department with DeptRunOptions::default_env(), supplying no per-run FKST_RUNTIME_LOG_DIR. fkst.codex_runs (live-run detection) therefore read the ambient user-global Codex log stream (~/Library/Logs/fkst) via sdk_codex's fallback. A matching running record — or a recent completed record from a preceding test invocation — makes a department (e.g. github-devloop-pr.fix) take skip-idempotent(live-exec-ref) after its driving delivery is accepted, so the graph quiesces without spawning the mocked Codex.

Effect: any run-graph test that asserts a Codex spawn is flaky — it fails whenever any ambient codex process runs on the host (e.g. concurrent operator workers) or a prior test seeded the global log dir, and passes in clean CI. This was diagnosed from a github-devloop-pr composed run-graph test (run_graph_polluted_merge_gate_stream_reaches_fix) that failed locally (ambient codex running) but was green in CI; controlled reproduction (seeding the global log dir) flips PASS→FAIL, and max_steps is irrelevant (a passing graph spawns Codex at step 5 of 11; cap 20 still fails with a seeded dir).

Fix (test-runtime only; production untouched)

run_graph now creates a per-invocation runtime_log_dir under the temp dir it already allocates and threads it through run_record to every department execution via FKST_RUNTIME_LOG_DIR (RUNTIME_LOG_DIR_ENV). One shared dir per run_graph call ⇒ codex-run statuses created within a graph stay visible to later steps of that same graph; a fresh dir per call ⇒ graphs are isolated from each other and from the ambient/global stream.

• Production sdk_codex (read_codex_status_records, runtime_log_dir/default_runtime_log_dir) has zero diff — real runs are unaffected.
• max_steps unchanged; no production live-run dedup change.

Test evidence (operator-verified outside sandbox)

• New regression run_graph_isolates_codex_logs_per_run_and_preserves_intra_graph_visibility: seeds a matching ambient codex-run record, asserts the in-graph spawn is NOT suppressed, and asserts intra-graph status visibility + inter-run isolation. RED without the fix (exited 101, "ambient codex status suppressed in-graph spawn", run_graph exceeded max_steps with pending=1); GREEN with it.
• Full run_graph_cli suite: 6 passed / 0 failed.
• cargo build -p fkst-framework: clean.

Review record

Operator diff-review: production isolation confirmed (no sdk_codex.rs diff), with_env_var helper is pub(crate) test-infra, per-graph tempdir scoping guarantees both isolation and intra-graph visibility, no existing test weakened. Diagnosed via an independent codex investigation (controlled flaky reproduction) as the sole root cause of the local-red/CI-green divergence.

🤖 Generated with Claude Code

https://claude.ai/code/session_01C8iuJ16FXCA7ejVruPEoZG

⟦AI:FKST⟧

[codecov]

Codecov Report

❌ Patch coverage is 0% with 9 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
crates/fkst-framework/src/test_runtime.rs	0.00%	5 Missing ⚠️
crates/fkst-framework/src/test_department_env.rs	0.00%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!