feat(handoff): dual-form context thresholds + model-preset registry

[DumoeDss]

Why

All context-occupancy thresholds (handoff default 0.5, reuse default 0.25, per-role overrides) are fractions of the context window. For 1M-context models a fraction works, but for small-window models (e.g. the GPT-5.6 family at ~300K) a fraction fires far too early in absolute-token terms — and those models degrade gracefully near their limit, so the remaining absolute headroom is the better signal. Configs need an absolute-token threshold form, and the CLI should ship sensible per-model presets so users of small-window models get good behavior without hand-tuning.

What Changes

• Handoff and reuse threshold values (pipeline-level, per-role, stage-level) accept a second form: { remainingTokens: <positive integer> } — an absolute required-headroom threshold — alongside the existing bare fraction in (0, 1]. A bare number is ALWAYS a fraction; the absolute form is ALWAYS the object, so no value is ambiguous.
• A built-in model-preset registry (src/core/model-presets.ts) keyed by model-id substring patterns provides, per known model family: the context-window size and suggested handoff/reuse threshold values. It subsumes the existing ad-hoc model→limit map in resolveModelLimit.
• Threshold resolution gains a preset layer just above built-in defaults: stage handoff > pipeline handoff.roles[<role>] > pipeline handoff > model preset (via the stage's resolved model) > built-in default. Reuse: reuse.roles[<role>] > reuse.threshold > model preset (via agents[<role>] model) > built-in default. Any ordinary config value therefore overrides a preset.
• rasen pipeline show --json reports resolved thresholds in whichever form they resolved to (bare number for fractions — byte-identical for existing configs — or the { remainingTokens } object), and source can now be preset.
• rasen agent context output gains remainingTokens (limit − contextTokens) so absolute thresholds can be compared directly against a probe.
• The orchestration playbook and handoff templates state the comparison rule for both forms (fraction: pct >= t hands off / pct <= t allows reuse; absolute: remainingTokens <= N hands off / remainingTokens >= N allows reuse).
• Backward compatible: existing fraction configs parse and resolve byte-for-byte identically; no version bump.

Review

Went through 3 rounds of an independent review-cycle loop — clean, with all Major findings resolved (test-tautology fixes for preset ordering and byte-identity guarantees, and a source provenance fix so pipeline show never mislabels a preset-sourced threshold as coming from a config layer). Full findings in the change's review-report.md.

Test plan

• Full suite: 134 files / 2796 tests, green (after merging in dev/0.1.4's concurrent fix/codex-host-compat merge and rebuilding dist/)
• tsc --noEmit clean
• rasen validate handoff-absolute-thresholds — valid
• Manual smoke: a fixture pipeline with handoff: { threshold: { remainingTokens: 60000 } } validates and shows correctly; a fixture with agents.implementer.model: gpt-5.x and no thresholds resolves source: preset

🤖 Generated with Claude Code

https://claude.ai/code/session_011an7Av8HWb8tWFrUPD9ZTi