Trace-derived journey ground truth (Phase 4 / B3)

[inth3shadows]

Replace hand-labeled journeys with real request traces, to measure the static selector's recall against reality instead of asserting it.

Both the parent plan (B3, "closes the loop the whole product is about") and TDAD's stated future work ("integrate dynamic coverage data to improve impact precision") point here — convergent evolution, so no moat, but it is the only thing that turns recall from a claim into a measurement.

Scored F3 x N4 = 12. Expensive; hold until #5 and #6 land. Effort: XL. Risk: High.

[inth3shadows]

Instrument shipped in #58; the measurement is still blocked

harness/trace.py + harness/plugin/tgtrace.py + harness/ground_truth.py are on main. They run a target's own pytest suite under sys.monitoring, record every function entered during each test body, and score traced(J) against couple.footprint(entries) — reporting traced_only (ran during J, cannot select J: the silent-miss shape), static_only (over-selection, ungated) and unresolved (no node at all — an indexing gap, not a traversal gap).

Validated end to end on harness/fixtures/dyndemo, where routes.create reaches dyn.audit only through getattr(mod, HOOK). The harness found it. That proves the instrument; it says nothing about honeyslate.

What remains, and the only thing blocking it: honeyslate's backend/tests/conftest.py declares clean_db as autouse=True and opens postgresql+psycopg://honeyslate@localhost:55433/honeyslate before every test. That port is closed on this machine, and Docker is not installed under WSL so deploy/docker-compose.yml cannot raise it. pytest --collect-only finishes in 0.05s; a real run blocks on connect and was killed at 300s. The suite is not slow — it is waiting for infrastructure that is not here.

harness/journey_tests_honeyslate.json is written and ready. The run needs a dev Postgres on :55433 and nothing else.

Worth weighing before standing that up: honeyslate's last commit is 2026-07-17 and signedintake's is 2026-07-19 — the only two repos with approved registries have been dormant for three weeks, which is also why #10's ledger will read zero. testgraph itself is the most actively pushed repo here and has a 238-test stdlib-only suite that runs in 7s with no infrastructure at all. Giving testgraph its own registry would be a cheaper trace target than reviving honeyslate's Postgres, at the cost of self-reference (the instrument and target share a repo — weaker evidence than honeyslate, though a trace is still independent of the graph it is scoring).

Full scope of this issue — traces replacing hand-labeled journeys — is untouched. One hand-labeled artifact survives by design, in its own visible file rather than as a heuristic inside ground_truth.py.

[inth3shadows]

Ran the self-trace proposed in the last comment: testgraph's own suite against its own registry, since honeyslate's Postgres is still unreachable. Full writeup in TECHNICAL.md Update 7, PR #67.

Confirmed the module-alias blind spot (Update 5) independently via trace, and found one new defect not explained by aliasing: ledger.append has zero real inbound static edges despite unaliased call syntax, while unrelated list.append() calls elsewhere false-positive-link to it — filed as #66. A commit breaking ledger.append currently selects neither J1 nor J5.

#12's full scope — traces replacing hand-labeled journeys — is still open.