perf(tests): premerge-findings fixture is ~96% interpreter startup #784

Surfaced by the Efficiency lens on #768 and deliberately not applied in #781, because it is not a drop-in swap.

One run of tests/premerge-findings.test.sh starts 374 python3 processes (counted with a PATH shim):

• 225 execute the module under test. That is the price of driving the shipped CLI with real argv and real exit codes, and it is not waste.
• 149 are scaffolding: building a small input document, writing a blocked-on artifact, reading one field out of a counts object, rendering a ledger. write_input alone is 81 call sites, write_blocked 11, counts_field 12.

Measured on an M-series host: the fixture takes 12.55 s wall. A bare interpreter start is ~19 ms (250 starts = 4.79 s) and a start that also imports the module is ~41 ms (20 starts = 0.83 s), so 225 × 41 ms + 149 × 19 ms ≈ 12.1 s — the fixture is roughly 96% interpreter startup, and the scaffolding half is ~2.8 s, or 22%.

#768 amplified an existing cost rather than creating it: its additions contribute 28 new write_input calls and 73 new driver calls, and per-section timing puts the added B32/B33 blocks at 3.89 s of the 12.55 s.

jq is already a hard dependency of this file — it exits FATAL at the top when jq is absent — and starts in ~3.4 ms (250 invocations = 0.85 s). Routing the scaffolding through it would take that 2.8 s to roughly 0.5 s.

Why this is not a mechanical swap

Each helper needs its own equivalence check:

• b32_growths asserts against Python repr spellings that jq renders differently.
• counts_field's missing-key sentinel needs a jq form that still separates a null value from an absent key.

Do this helper by helper, verifying every assertion string each one feeds. A blanket swap will turn assertions green against different text.

Do not extend it to the driver calls

The 225 module executions are the fixture's whole value — they drive the shipped CLI rather than a transcription of it. Replacing any of those with an in-process import would re-introduce exactly the replica-test class this suite exists to refuse.

Related: PR #768, PR #781