refactor(do): remove fable from the model selection policy

[notque]

Directive

Owner, verbatim: "our document for what agent to use should remove fable entirely". PR #891 demoted Fable-5 to a labelled "prior measurement" row and a manual-only pick. This removes it.

What was removed

skills/meta/do/SKILL.md — the Fable-5 variant-table row, the manual-pick set and its ordering clause, the openai max-power miss → fable/xhigh escalation target, and the felt-quality axis clause fable > sol. Sentences that depended on fable were rewritten, not left dangling (the manual-only rationale moved from "felt-quality escalation" to the constraints the remaining rows actually serve).

Code and other live routing instructions: scripts/build-dispatch.py (VALID_MODELS, ANTHROPIC_MODELS, the ("fable","sonnet") manual gate → sonnet), hooks/routing-decision-recorder.py valid-model set, scripts/validate_component_models.py ALLOWED_MODELS, docs/PHILOSOPHY.md Model Policy by Task Class, skills/meta/agent-creator/references/agent-frontmatter-template.md, agents/system-upgrade-engineer/references/component-audit-checklists.md. Tests updated to match, including the SUPPLIED_CLAUDE_POINTS fixtures and the fable-only cost-monotonicity test (re-pointed at sonnet, which still exercises the same property).

No agent or skill frontmatter declared model: fable, so nothing needed migration.

What was kept, and why — Opus-4.8 and Sonnet-5

The directive named fable only, so this is a judgment call, stated explicitly rather than made silently: both remaining prior-measurement rows stay, as manual-only picks.

1. They are the Anthropic lane's only DeepSWE data. Opus 5 is unbenchmarked (n/a in every cell), so deleting them leaves the lane with no numbers at all and no basis for the start-low-escalate-on-miss cost argument.
2. They answer constraints the benchmark does not measure and Opus 5 does not cover: cost step-down, latency, context window, fan-out breadth. Fable's rationale was felt quality — that rationale left with fable; theirs did not.

If the owner wants the Anthropic lane reduced to Opus 5 alone, that is a one-line follow-up, but it is a different decision from this one.

Kept as historical record

docs/what-didnt-work.md (dated A/B entry naming fable arms and judges) and scripts/routing-ab-results/ (model-ab-v1 PROTOCOL/REPORT, self-route-v1 VERDICT). Editing these would falsify dated experiment records, and neither is a live routing instruction — they describe what was run, not what to run. skills/review/security-review/references/coverage.md matches the grep only on the word "Spoofable"; out of scope.

Also out of scope: PR #891's component list named skills/workflow/references/workflow-patterns.md and skills/meta/codex/SKILL.md as fable sites. Neither contains a fable reference on main — noted as a correction in the ADR rather than acted on.

Traps checked

• No dollar-digit (commit 758b862): grep -n '\$[0-9]' skills/meta/do/SKILL.md → empty. All costs remain plain numbers.
• Off-policy invariant: manual picks still require manual_model_override=true and model_effort. The gate set narrows from ("fable","sonnet") to sonnet; the requirement is unchanged and still tested.

Gates

ruff check + ruff format --check clean; pytest scripts/tests/ 4789 passed, 1 skipped, 150 xfailed; hooks/tests/test_routing_decision_recorder.py green; scripts/validate-do-references.py OK (55 mentions resolve).

ADR adr/model-selection-policy.md amended locally (gitignored) with the directive, the retention reasoning, and the correction above.

[github-actions]

Skill-eval coverage

Changed skills in this PR: 1 (426ef925b7f3..a7dc37157144).

Report only. CI cannot run evals (the runner needs the claude CLI). Run the ablation locally to get the base->head delta.

Mapped

• do -> eval evals/new-skills-ab-test. Run locally: make skill-eval-ablation BASE=426ef925b7f3 HEAD=a7dc37157144 SKILL=do