Implement SPEC-09: relaxation / least-worst under infeasibility (Stage 4)

[IanMayo]

What & why

Stage 4 of the build plan — least-worst (thesis B live). POST /relax turns an infeasible commitment set into an argument the commander weighs, never an error the machine hides or a scalar it optimises. On the Meridian R3m world (both approaches mined, causeway dropped) no plan satisfies C2, C3 and C4 together; /relax returns the three inclusion-minimal least-worst ways forward, each naming what it gives up.

Research-first (DEC-11 gate)

docs/research/04-relaxation.md is authored before implementation and decides:

• Formalism = preemptive lexicographic priority over the must/should/prefer tiers, NO numeric weights. Weighted CSP, partial weighted MAX-SAT, and Archimedean goal programming are rejected on honesty grounds (DEC-19), not scale — a numeric sacrifice weight is a ratio the commander never issued. This spends ASSAY-FIND-2 at the build face.
• Return the inclusion-minimal correction-set frontier, not one optimum (the MCS view of infeasibility) — ranked least-worst-first, the must-sacrifice kept but ranked last (G4: never silence).
• sacrificed = exactly the violated commitments, grounded in the four-stop scale the reused scorer computes; tight/marginal are risks, not sacrifices.
• Content-neutral, stated tie-break for same-tier sacrifices (C3/C4 are both should) — because encoding a civil-harm-vs-force-protection ranking into code is the value judgement DEC-19 exists to prevent, in non-numeric clothing.
• The note also corrects course honestly: the SPEC-08 BASE generator routes clear of the banded regions and cannot enter the mined water, so SPEC-09 authors an R3m-responsive candidate set (route latitude, vignette §8) while the sacrifices stay computed by the scorer (DEC-10).

Changes

Specified via spec-kit (specs/009-relaxation/) and implemented:

• src/tiers.ts — tier lexicography + inclusion-minimality over sacrifice sets + the stated tie-break (pure, no weights)
• src/relaxCandidates.ts — authored R3m-responsive candidate plans (routes documented cell-by-cell)
• src/relax.ts — RelaxService (/relax): generate → score (reused ScoreService) → read violated sacrifice sets → inclusion-minimal → rank by tier → cited_in/sacrificed_in edges
• src/components/s3Cards.ts — the S3 least-worst cards
• new seam movement types; no schema change (RelaxationReport/RelaxationCandidate/Plan already exist)

Stage-4 exit (verified on Meridian)

/relax over R3m returns three inclusion-minimal least-worst candidates sacrificing C4, C3, C2 — each sacrificed non-empty and computed (re-scoring each plan violates exactly its set; the {C2,C4}/{C3,C4} supersets are inclusion-dropped), the C2 must-sacrifice ranked last but present, the C3/C4 same-tier tie stated in tie_break, every card in command language and traced back via cited_in/sacrificed_in (G3), never empty and never a silent drop (G4).

Invariants

G4 (least-worst, never silence) upheld by the report shape and validation (empty candidates/sacrificed fail validation); G2 (no decimal/verdict-token on the card face; margin in the trace drawer); G3 (complete backward chains); G1 (deterministic over (world stamp, commitment ids, seed, engine)); DEC-19 (ordinal tiers, no weights, stated tie-break).

Testing

• npm test — 187 pass (163 baseline + 24 new: 13 tiers, 11 relax)
• npm run typecheck — clean
• npm run gallery — the "least-worst, never silence" demo moment renders the three R3m cards from the actual /relax

Register & propagation

No new register decision — the formalism is decided by DEC-19, the report shape by knowledge model §5/§8 and seam §7/§G (G4), the tie-break rule by DEC-19's own "state it explicitly." Batch-propagated in the same change: CLAUDE.md current-phase line, delivery-plan status, gallery.

🤖 Generated with Claude Code

---

Generated by Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-13 18:55 UTC