fix(benchmarks): resolve a tag to Ollama's own spelling, and mark a model whose every run failed

[Xore]

Two defects that combined into silent data loss. Three models pulled
perfectly, failed all six run attempts, were written MODEL_DONE, and had their
weights deleted — leaving nothing in the results directory to say they had not
been measured.

1. Ollama's stored tag case is not the case you asked for

It rewrites some quant-shaped tags and not others. Measured on tags this repo
created itself, every one written lowercase:

gemma4-26b-a4b-selfquant:q3_k_m    ← kept as written
gemma4-26b-a4b-selfquant:Q4_K_M    ← uppercased by Ollama
gemma4-26b-a4b-selfquant:q5_k_m    ← kept as written

sweep_extra.sh's presence check is grep -qixF — case-insensitive, so a
folded tag passes. record_baseline.py then resolves the model by exact
string match against /api/tags and exits model tag is not installed. Every
run fails for a model that is sitting right there.

It hit exactly the rows that matter most: Ornith-35B Q3_K_M and
gemma-4-26B-A4B Q3_K_M/Q5_K_M — the published twins of the self-quantized
ladder, i.e. the controls the entire #2245 comparison rests on. gemma Q4_K_M
succeeded, which is what made it look like a per-model quirk instead of a
systematic one.

Fixed in the driver, not in record_baseline.py — that file is on the pinned
a99e765 harness, and moving the pin to fix a name lookup would split the
scoring vintage for no benefit. The driver now hands the scorer the name Ollama
actually holds.

Verified live: the same model that failed three times now scores 62.

08:31:01 tag case differs: roster '…:Q3_K_M' -> ollama '…'
08:32:49 done  A Ornith-1.0-35B…Q3_K_M run1 score=62

2. A model that produced nothing left no marker

Total failure wrote a GIVEUP line to failures.txt and MODEL_DONE to the
log, and nothing next to the results. Anything reading the results directory
could not distinguish it from a model that was never in the roster. Same
silent-hole class as the false EXTRA_COMPLETE in #3031 — and it is why defect 1
went unnoticed until the phase-3 chain refused to advance.

Now writes UNMEASURED with the reason, so the state is machine-readable.

The guard held, and that is the point

chain_phase3.sh's completion check — every roster entry has both tier files
or an explicit marker — blocked with:

sweep not running but roster incomplete (46+3/52) -- waiting

Without it, the ladder scoring and then the 96-model cold run would both have
proceeded with three of the most decision-relevant rows missing, and nothing
announcing it. The whole reason that guard exists is that EXTRA_COMPLETE has
lied three times; this is the fourth shape it would have taken.

Refs #1947, #2245, #3031, #2738

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None