test(aryabhatta): account for every attempt the calibration run samples

[adityak74]

The 60-directory registry run reported unusable 25, but only 19 lines in the
log said unusable. Six attempts looked like silent losses, and the summary
figure was computed as sampled-minus-scored, so it counted them without knowing
what they were.

They were not lost. I was looking for the wrong word.

Three discard paths printed three different phrasings and the summary borrowed
the word from only one of them. The six are all there, under no answer (...):

4  no answer (https...)                      network errors reaching the model
1  no answer (Connection reset by peer ...)  os error 54
1  no answer (step limit reached)

So the harness was not dropping attempts. It was describing them in a
vocabulary that did not match its own total, which is worse in one specific
way: the count looked checkable and was not.

What changes

Every sampled directory is now scored or discarded into one of eight named
categories. A Tally counts and prints in the same call, so neither can
happen without the other, and the run asserts that scored plus discarded
equals sampled before it prints a calibration report.

The per-attempt work moved into record_attempt, which takes the store write
as a closure. That is what lets the whole accounting path run in tests with no
model and no database.

Categories report zero rather than disappearing. PR #86 fixed both parser
reasons, so no fenced json block and not the shape asked for should now
fall to zero. A category that vanishes when it stops firing takes the evidence
of the fix with it.

A discard is never a way to raise the pass rate

This is the line that matters on a change to a measurement harness. Nothing
invents an interval, and an unreadable answer contributes nothing to n.
a_discarded_attempt_never_becomes_a_scored_one exists to say so.

Discards now carry the text that failed

The two parser bugs #86 fixed could not be replayed, because only the error
survived and the model's actual answer was never recorded. A discard now
carries an excerpt of the answer next to the reason.

That text is model-authored. It goes to stdout and nowhere else. It is never
written to the store, so no detector and nothing in the search layer can read
it back, which is the integrity rule in CLAUDE.md.

Why this matters for the go/no-go

n clearing MIN_CALIBRATION_N = 50 is the current blocker on getting a real
verdict instead of NotEnoughEvidence. Six unexplained losses per sixty is
directly material to that, and a harness whose own total cannot be reconciled
against its own output undermines every number downstream of it.

Verification

cargo test -p zorp-agent --features research --test evidence_calibration
16 passed, 0 failed, 1 ignored (the live one). Among them
the_registry_run_that_seemed_to_lose_six_attempts_reconciles, which
reconstructs this exact discrepancy.

cargo fmt --all --check clean.
cargo clippy --workspace --exclude zorp-track --all-targets --locked -- -D warnings clean.

Provenance

Written by a subagent that committed and pushed, then died on an API session
limit before it could open the PR or report results. I verified the branch
myself: I confirmed the six no answer lines in the original log, ran the
tests, and ran fmt and clippy. Its own final coherence read never happened,
so treat CI and review as the first full check.

https://claude.ai/code/session_01KGPVQ8wUG7h36zashWYCp4