Faber spawns coders with configured model + bounce protocol (no runtime escalation)

[yihanzhu]

Context

Coder subagents currently inherit Faber's frontier session model — the dominant usage-throttle driver. Approved design: coder runs at a fixed sonnet ceiling; capability ceilings are load-bearing (they force scope discipline upstream). Depends on the config/models.conf foundation issue.

Spec

• manager/CLAUDE.md + templates/faber-command.md: at every coder spawn, Faber reads config/models.conf (plus the target repo's override file if present) and passes an explicit model parameter = FABRICA_CODER_MODEL on the subagent spawn. Never escalate the tier at runtime.
• Bounce protocol (replaces any notion of model escalation). On a bounced review round Faber diagnoses:
  1. Spec gap → amend the revision brief with a Faber-authored diagnosis; respawn at the same tier.
  2. Scope too big / genuinely hard → decompose: land the independently-green mergeable core; file AND link the follow-up issue BEFORE merging the partial PR; a follow-up inherits the parent's approval only as a strict subset of the approved scope.
  3. Stuck / reviewer disagreement → rounds cap → needs-human, unchanged. Decomposition happens within the cap and never extends it.
• Persona: Faber's own session is expected to be frontier-tier; if Faber detects otherwise, warn the user.

Sync note

Prompt-only change — takes effect ONLY after re-running scripts/install.sh to regenerate the live /faber command. Call this out in the PR description.