TDD 0066: /build-tdds model dispatch, confirmation, and run record

[cahenesy]

Implements TDD 0066 (PRD f6ef178, ADR 0015). Stacked on #184 (0065). Once #184 merges, the base retargets to master automatically, since auto-delete is on.

What changed

• models.sh: tl_dispatch_model_arg, tl_model_warnings and tl_models_confirm.
  • The queue confirmation shows the parent model, the effort and its source, and each worker's model and source.
  • Two warnings: one when a judgment slot is pinned to the light tier, one when an effort pin is ignored.
• run-record.sh: a per-TDD sidecar <run>/<slug>.models.json with 17 keys. It is written read-overlay-write, atomically, with JSON-escaped values. tl_run_set_tdd and tl_run_set_pr never touch it.
• skills/implement/SKILL.md:
  • Block contract: every block is self-contained and takes its inputs from ${TL_*:?} env vars.
  • Two new blocks: tl:models-confirm (step 5) and tl:models-record (step 7).
  • Dispatch rule: inherit means no model parameter.
  • Step 9: resolves from the TDD path.
  • Step 10: "prior-gen" removed.
  • Logging: each worker logs its model.
• README.md: the models paragraph is rewritten, and every "different model" claim now says "fresh worker".
• tests/model-dispatch.test.sh (136 checks) runs the functions and the extracted blocks with env vars only.
• Version 3.47.0 → 3.48.0, plus a releases.json entry.

Gates

Gate	Result
test-first	PASS (de0bee0 red: 76 fail → 9be7de9 feat)
ci-checks	PASS (8 evals; model-dispatch 136/0)
runtime-verify (fable)	PASS: 13/13. The real confirmation block on this session printed parent=claude-opus-5-5 effort=xhigh (settings; session-wide)
review (opus)	PASS: no blockers or majors

Review follow-ups (not blocking)

• Minor (0064 design gap):
  • When it happens: CLAUDE_CODE_SUBAGENT_MODEL is set and the parent model is light-tier or unreadable.
  • What goes wrong: mechanical verify is recorded as inherit [parent-cap] but runs on the env model.
  • Fix: one resolver arm that treats it as a pin, plus an eval case. I'll carry this into 0067's build.
• Minor: the step 9/10 model log lines are written by the parent from prose instructions. The eval only checks their text.
• Minor: a resume re-runs the record block, which logs implementer model= even when the implementer is skipped.
• Nits:
  • README and SKILL text say the env pin applies to "every slot".
  • status.sh skips the sidecar only by accident (it has no slug key).
  • effort_source is stored as - when the effort is unknown.

Merge note

Merge #184 first with a regular merge, not squash, then this one. Once this merges, claude plugin update is safe.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Ki3TDnP2bNqqpEkFoZaEXq