fix(decide,llm): close review gaps in model detection and completion advice

Review of the #238 -> #243 stack found three accuracy gaps in behavior that
1.7.9 would ship for the first time:

- `claude-mythos-preview` (and any unversioned Fable/Mythos id) did not match
  the model-trait pattern, so it still received `temperature` and got an
  HTTP 400. Fable/Mythos ids now match with or without a numeric version.
- A remote `verify_completion` probability between the certainty bound (0.75)
  and the 0.85 completion bar was reported as a decisive `is_complete=false`.
  It is now `uncertain` with a null conclusion, matching the documented
  "uncertain completion stays null" contract.
- The local completion heuristic missed Mocha/Jest-style `N passing`, treated
  `errors: none` as a failure, and needed `not passing`/`0 passing` handling
  once `passing` counts as success.

Every new test fails on the previous head. Rebind the public offline fixtures
to immutable v125 evidence; aggregates are unchanged from v124.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014ksPGXPZQTikJYpaGHa6R6

Author: claude

BENCHMARKS.md

@@ -94,14 +94,14 @@ interpretation and do not count as additional benchmark-quality gains.
 ### Public numeric evidence registry
 
 Every exact public aggregate retained below comes from the checked-in, public-safe
-[`offline-fixtures-v124.json`](docs/benchmark-evidence/offline-fixtures-v124.json) artifact. Its
+[`offline-fixtures-v125.json`](docs/benchmark-evidence/offline-fixtures-v125.json) artifact. Its
 SHA-256 is
-`dc9e35cc4c20f164d6adb33858bd444f8001f4f9dad7ad1be618e8c567adbcf6`, also recorded in the
+`1f74971d6213a188b31cf58f6ff6132a487da36d22455292b028dadc202a8feb`, also recorded in the
 adjacent `.sha256` file. The artifact contains no raw questions, answers, prompts, customer data,
 or per-record content fingerprints.
 
 The fixture-suite digest is
-`05007294ab0c835fea98d65239576f3221084213e2e022d8c84328e12014a6cc`. The artifact defines
+`42cb269b867a1e3210d0da77d6a040abc974cd44317b91ad10a2d483d07c1586`. The artifact defines
 the digest algorithm and records the SHA-256 of every suite and dataset file. Each evidence ID
 also binds its exact command through `sha256(UTF-8 exact command)`:
 
@@ -123,10 +123,10 @@ Historical LoCoMo, graph, handoff, consolidation, and security figures remain pr
 source artifacts but are omitted from the current chart until each has a matching immutable,
 public-safe artifact. The chart labels coding outcomes, external datasets, and operational
 capacity as pending evaluation tracks rather than implying scores. Regenerate it with
-`python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-v124.json --output docs/images/context-efficiency.svg` after selecting the report to publish.
+`python scripts/render_benchmark_report.py --report docs/benchmark-evidence/offline-fixtures-v125.json --output docs/images/context-efficiency.svg` after selecting the report to publish.
 
 The companion examples are also generated from that artifact with
-`python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-v124.json --output docs/images/evidence-backed-agent-examples.svg`.
+`python -m scripts.render_benchmark_examples --report docs/benchmark-evidence/offline-fixtures-v125.json --output docs/images/evidence-backed-agent-examples.svg`.
 The historical-to-executable mapping is in
 [`docs/BENCHMARK_CHANGE_COVERAGE.md`](docs/BENCHMARK_CHANGE_COVERAGE.md).
 

CHANGELOG.md

@@ -33,7 +33,13 @@ All notable changes to Engraphis are documented here. Format loosely follows
   the API defaults, `.env.example`, and the provider guide with `claude-sonnet-5-5`.
 - Documented how to back the experimental Jev decision adapter with Claude: pin an exact model
   id, keep fallback disabled, and avoid sampling parameters and forced tool choice.
-- Refreshed the public offline fixtures and source bindings in immutable v124 evidence.
+- Recognized `claude-mythos-preview` and other unversioned Fable/Mythos ids, so they no longer
+  receive `temperature` (an HTTP 400) and get the thinking-model output headroom.
+- A remote `verify_completion` probability between the certainty bound and the 0.85 completion
+  bar is now `uncertain` with a null `is_complete`, instead of a decisive failure.
+- Local completion checks recognize `N passing` (Mocha/Jest style), `not passing`, `0 passing`
+  and `errors: none`.
+- Refreshed the public offline fixtures and source bindings in immutable v125 evidence.
 
 ## [1.7.9] - 2026-09-29
 

README.md

@@ -84,9 +84,9 @@ neither is an end-to-end question-answer score. Coding outcomes, external datase
 operational capacity remain separate pending evaluation tracks until their artifacts are selected.
 
 These values are evidence IDs `offline-chunking` and `offline-performance` in
-[`offline-fixtures-v124.json`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/docs/benchmark-evidence/offline-fixtures-v124.json),
+[`offline-fixtures-v125.json`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/docs/benchmark-evidence/offline-fixtures-v125.json),
 SHA-256
-`dc9e35cc4c20f164d6adb33858bd444f8001f4f9dad7ad1be618e8c567adbcf6`.
+`1f74971d6213a188b31cf58f6ff6132a487da36d22455292b028dadc202a8feb`.
 [`BENCHMARKS.md`](https://github.com/Coding-Dev-Tools/engraphis/blob/main/BENCHMARKS.md#public-numeric-evidence-registry)
 records the matching suite digest, exact commands, and per-command config digests. The offline
 fixture registry intentionally excludes external, model-dependent, consolidation, productivity,

docs/LLM_PROVIDERS.md

@@ -109,9 +109,9 @@ Choosing a model:
 - Retired ids such as `claude-3-5-sonnet-20241022` and `claude-3-5-haiku-20241022` are rejected by
   the API; the connection test reports them as an HTTP 404.
 
-Opus 4.7 and later, Sonnet 5 and later, and Fable reject `temperature` and similar sampling
-parameters, so Engraphis omits them for those models. Opus 5 and later, Sonnet 5 and later, and
-Fable also think before answering by default. Engraphis sends `ENGRAPHIS_LLM_EFFORT` (`low`,
+Opus 4.7 and later, Sonnet 5 and later, Fable and Mythos (including `claude-mythos-preview`)
+reject `temperature` and similar sampling parameters, so Engraphis omits them for those models.
+Opus 5 and later, Sonnet 5 and later, Fable and Mythos also think before answering by default. Engraphis sends `ENGRAPHIS_LLM_EFFORT` (`low`,
 `medium`, `high`, `xhigh`, or `max`; default `medium`) for those models and keeps at least 4096
 output tokens available so hidden reasoning cannot crowd out the reply. Other providers and older
 Claude models ignore the setting.