fix(agent): a stop that read nothing is answered for whoever stopped

[yusuf-gundogdu]

The drive holds a sentence written for exactly one ending:

Read it yourself. Call inspect_schema for the tables and their columns, and inspect_plan for
how a statement will run, then call compose_report with what you found.

It was sent only to a model whose profile asked for it. One profile of twenty-eight does. A
model nobody has measured has no profile at all, so the model most in need of the sentence was the
one guaranteed not to receive it.

What the ledgers say

Measured across the sweep behind 0.14.1:

	runs
ended model-stopped with no-report	300
of those, having called no tool at all	150
of those, ended with the drive saying nothing	94

By model: mistral-small3.1:24b 51 · granite3.3:8b 41 · mistral:7b 40 · gpt-oss:20b 18.
None of the four is on the shipped roster — they are exactly the unmeasured models the sentence
could not reach. A hundred and ten of the hundred and fifty were on investigation, the first
surface a model meets.

Why the bound was protecting nothing

The gate's own comment had already argued the cost away. It fires only where anyToolCalled is
false, and compose_report is one of the tools that counts, so a run reaching it composed no report
and has already earned no-report — in the comment's words, "the turn cannot cost a pass."

A bound that cannot protect a passing run is not protecting anything. What this one withheld was a
sentence on a run already lost.

The change

investigation.ts now reads answersUnreadStop rather than retriesUnreadStop. The two differ on
one thing:

profile says	before	after
retryUnreadStop: true	sentence sent	sentence sent
retryUnreadStop: false	not sent	not sent — a measurement, obeyed
no profile at all	not sent	sent — an absence is not a measurement

This is the same split planStatementAsksFor made beside planStatementRetriesFor, and the fourth
setting found in this shape. All twenty-eight shipped entries state the field, so no shipped
model's turn count moves — verified against measured-profiles.json rather than assumed.

Nineteen red tests, which is the point

Nineteen existing tests turned red, and every one of them is a run that stops having called
nothing — the change reaching exactly what it was aimed at rather than spilling past it. Each gets
the extra scripted turn the drive now spends. tests/evals/real-model.ts holds the shared warm-up
that accounted for several of them at once.

The pair that pins the new behaviour lives beside the tests for the gate it belongs to, replacing
the one that asserted the old contract. Both halves were confirmed to bite:

Mutation	Result
an absent profile read as false again	1 fail
a stated false ignored	1 fail

Docs

docs/llms/model-tuning.md gains the section this makes necessary: the last column of the settings
table is the compiled default, and two settings answer differently for a model with no entry at
all. docs/AGENT.md, docs/BACKLOG.md B65 and the notes on profile.ts and models/index.ts are
corrected where they said the absent case resolves to the defaults.

Gates

Every step of the Lint, Typecheck and Build job, in order: chart:check (strict) ·
channels:showcase:check · readme:check · security:check · format · lint · typecheck ·
knip · build · build:lib · attw — all green. test:coverage and plain test both exit 0.
Coverage: 45995/45995 lines, 100.00%.

Rebased onto main after #633 merged; no file here is touched by any other open PR.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud