chore(ci): count review denials from every source, and census the run

[thecodedrift]

The #185 guard caught the zero-turn review on #196, which is what it exists for. Then it reported the wrong reason.

What went wrong

The guard printed permission_denials=0. The streamed job log for the same run carried permission_denials_count: 6.

That zero was then used as evidence that permissions were not the cause, which sent the investigation the wrong way. Reporting the wrong number was worse than reporting nothing, because it looked like an answer.

Measured on run 33039882331: the saved execution file's result record omits permission_denials_count, while the streamed stdout carries it. Every other field matched exactly, num_turns, is_error, and total_cost_usd to sixteen digits, so the guard found the right record. The file just lacks that one field.

The naming loop was already there. It sat behind if denials:, so a missing count made it unreachable:

denials = result.get("permission_denials_count") or 0   # 0 when the file omits it
...
if denials:                                             # never true
    records = result.get("permission_denials")
    if not isinstance(records, list):
        records = [m for m in messages if "denial" in ...]   # the scan that would have worked

What changes

Denials are counted from every source. The maximum of the count field and the number of denial records found in the file, and the naming loop reads those records directly instead of re-deriving them inside a gate that could not open.

The run now describes itself. A structural census of the execution file: message types and counts. Two zero-turn runs have billed real inference, #182 at $1.08 and #196 at $2.20, and neither left anything to diagnose from afterwards.

Types and counts only. No message content, no tool inputs, nothing the model produced while reading an untrusted diff, which is why show_full_output stays off on a public repo. None of that is needed to say what shape a run had.

Verification

Three fixtures, covering both observed shapes:

fixture	before	after
#196 shape: count omitted, denial records present	0 denials, nothing named	2 denials, names TodoWrite, Task
#182 shape: count present	4 denials	4 denials, unchanged
healthy run	exit 0	exit 0, unchanged

All three also emit the census line.

What this still does not tell us

Why the run takes zero turns. Six denials on #196 and four on #182 make the allowlist the strongest hypothesis again, but a denial does not have to be fatal, and nothing here explains why inference is billed before any turn completes. The census is what should narrow that next time: the shape of a zero-turn run is currently unknown, and one run of it will now say so in the log.

Refs #185