fix: honeyslate's registry declared no blind spots, and had two

[inth3shadows]

Closes #62.

Found by the same structural sweep that found testgraph's own blind spot in #59: seed every node of each product file, run the impact closure, intersect with the entry map. 17 of 38 honeyslate product files reached no journey. Most were expected — 6 alembic migrations, 6 frontend files TECHNICAL.md already excludes. Two were not.

The two real gaps

file	why it is a real gap
app/main.py	lifespan calls scheduler.start(). Break it and the auto-scheduler never runs — yet nothing selected J8 for a change to it. A recall bug, not a registration preference: the journey existed, its entries were just one hop short of the thing that starts it.
app/routers/google.py	a live wired router (app.include_router(google.router)) serving /api/google/status and /selftest, with no journey at all

Closed by extending the entries of existing journeys, deliberately not by adding journeys — lifespan/create_app join J8, the two Google diagnostics join J7 (same integration as gcal_push). A J9 would have changed the registry's journey set, which is what labels_honeyslate.json and the published "mean 3.33 of 8 journeys named" are both written against.

app/cli.py stays uncovered and is now declared, along with the migrations and the frontend.

Why an undeclared gap here costs more than elsewhere

The registry previously declared nothing — and because it is approved: true, the UNAPPROVED REGISTRY banner whose entire job is to say "a NONE may mean not-registered rather than not-affected" never fired for it. honeyslate is the only repo whose numbers this project quotes as accuracy rather than sizes.

Re-measured, because the registry changed

harness/accuracy.py reads journeys/honeyslate.json live, so any edit silently re-points every published number at a registry that no longer exists — the exact defect 94e29a9 corrected for signedintake. Rather than assert that four added entries were harmless, both registries went through the harness back to back:

registry	entries	min recall	mean precision (non-empty oracles)
before	16	1.00	0.68
after	20	1.00	0.68

S1 PASS either way.

The honest reason precision did not move: the five labeled commits barely touch the newly covered files. This shows the entries cost nothing on the scored set, not that widening entries is free in general. A future commit touching app/main.py will now select J8 — which is the point of the change, and exactly the case the labeled set does not contain.

Verification

entry nodes: 20 | unresolved: none
  backend/app/main.py            -> ['J8']
  backend/app/routers/google.py  -> ['J7']
  backend/app/cli.py             -> NONE   (declared)

All 8 journeys resolve, no unchecked_entries, no approval warning. 251 tests green.