docs(causal): accept DESIGN-causal.md — typed causal hypothesis graphs (review-only)

[imaimai17468]

Problem (issue #320)

Phases #321–#323, #360, #364 of the causal profile all cite an accepted docs/DESIGN-causal.md as their sole design authority, but the normative content lived only in issue #320's body, with eight open decisions unresolved.

Change

Add docs/DESIGN-causal.md as the accepted design contract. It preserves #320's normative content — three-plane architecture, guarantee-boundary table, surface syntax + variable/claim field contracts, scope declaration & three-valued containment, claim content version & retired lifecycle, time semantics (lag/persists, Minkowski sums, unknown propagation, fail-closed clock mapping shared verbatim with #323), feedback semantics (instantaneous-loop error, polarity sign products, unknown absorption, loop_class witnesses, the narrow feedback_without_damping_story rule), typed CausalModel IR + invariants, JSON contracts, authority surface, and algorithm constraints — and resolves the eight open decisions (§10):

#	Decision
1	standalone causal block is final; nested sugar not adopted
2	timebase = closed enum tick|hour|day|week; custom units rejected in v0
3	persists adopted (required claim field); no quantitative convolution
4	uses <alias> from "<path>" + alias-qualified 2-segment refs
5	undeclared positive-lag cycle = warning (matches #321's causal_unacknowledged_feedback); instantaneous loop stays error
6	evidence-artifact validation confined to Phase 2; Phase 1 never opens artifact files
7	causal schemas stay independent under schemas/fslc/causal/
8	moderated_by deferred — reserved, no v0 grammar

Also: docs/README.md design map row, CHANGELOG.md [Unreleased] entry.

Acceptance criteria mapping

• feat(causal): causal check/analyzeとcausal-graph.v0を追加する #321 can start from this document alone: syntax, IR, diagnostics severity, JSON shapes, and algorithm constraints are all normative here.
• "No changes to Public Kernel v1 / runtime / solver / verifier" is stated as a design invariant (§1, §9).
• No output path attaches proved/verified to a causal claim (§1, §11 non-goals).
• Scope three-valued judgment, polarity products/unknown absorption, and loop classification are normative with Phase 1 test obligations pinned (§12).
• Claim version pinning + retired lifecycle guarantee old-version evidence never supports a current claim and refuted history is never lost (§4.5).
• Clock-mapping fail-closed conditions are identical to feat(causal-expectation): 明示的な観測期待を既存kernel propertyへ限定loweringする #323's expectation-compiler gates (§5).

Notes

• Docs-only change; no Rust/schema/CLI surface moves in this PR.
• Prerequisite spike spike(causal): 長期遅延・持続・フィードバックを含む因果仮説プロファイルの需要と意味論を検証する #319 is still open; proceeding was an explicit maintainer instruction (treated as the go decision). If spike(causal): 長期遅延・持続・フィードバックを含む因果仮説プロファイルの需要と意味論を検証する #319 concludes no-go, this design can be marked superseded without code impact.

Closes #320

🤖 Generated with Claude Code

[rizumita]

Ran an independent adversarial design review of the full 734-line docs/DESIGN-causal.md. Phase 1 (check/analyze/diff) is specified well enough that two implementers would converge — good work there. But given this doc is being accepted as the sole normative authority for a 5-phase implementation (PR #369 already ships all 4 phases against it), a few gaps matter enough to request changes before final acceptance:

1. Claim-version/retired lifecycle (§4.5/§12) is honor-system. No content-digest pinning is specified, so a claim's content (e.g. polarity) can change without a version bump while old evidence keeps counting as current support. retired is never stated to be terminal — a claim could go retired → active again with stale evidence silently re-activating. "Re-proposing an already-refuted hypothesis is detectable" names no actual mechanism.
2. Authority-scope mismatch for Phases 2–5. The evidence schema/lifecycle (#322), support-aggregation table, and observation bridge (#360) are deferred to bare issue references with almost no normative content here. The observation bridge in particular (production replay → evidence feeding causal_support) looks like exactly the "temporal co-occurrence becoming causal support" the doc's own §11 lists as a non-goal, with no stated constraint distinguishing the two. Similarly, the expectation compiler's leadsTo lowering (§3, real kernel-provable) has no output-shape contract for how a proved verdict on the lowered form should be labeled relative to the causal claim — exactly the seam where a reader could misattribute validated-causality semantics.
3. confidence: 0.92 in the §8 JSON example has no defined semantics, contradicting the doc's own no-probability/fully-deterministic stance.
4. basis: supported vs. the content-version contract: unclear whether derived support is version-relevant content or presentation-only — implementers could reasonably build either, and one choice is a version-integrity bug.
5. Feedback-cycle classification with lag: unknown edges (§5ק6 seam) is undefined for a cycle with all-zero known lags plus an unknown-lag edge — could misclassify an instantaneous (physically impossible) loop as an allowed delayed one, or vice versa. (Note: I found this materialized as an actual bug in PR #369's implementation — rust/fsl-tools/src/causal.rs's validate_loops excludes Lag::Unknown from the zero-lag check, contradicting its own doc comment.)
6. Evidence-side scope-dimension comparison (dimension present on one side, absent on the other) is unspecified.

Recommend either specifying sections for the #322/#360 contracts here, or explicitly scoping this doc's "sole authority" claim to Phase 1, plus tightening 1–4 above. Happy to re-review once addressed.

[imaimai17468]

@rizumita レビューありがとうございます。6件すべて対応しました（PR #369のブランチで修正、このPRにマージ時に反映されます）。

1. Claim version/retired lifecycle (§4.5)

• content変更をversion bump無しで行った場合、causal diffがcontent_changed_without_version_bumpとして検出
• retiredはterminalと明記 — retired→active再遷移は不可、再提案には新IDが必要
• 再提案検出: causal diffがretired claimと同じsource/target/polarityの新claimをretired_hypothesis_reproposedとして報告

2. Authority scope (§1)

• §1にフェーズごとの規範的権威範囲を明記
• observation bridge (§16) はtemporal co-occurrenceをdirected supportに変換しないことを明示
• expectation verdict (§14) のラベルはkernel propertyに付き、causal claimには付かないことを明示

3. confidence: 0.92 (§8)

• deterministic structural score（グラフ構造から計算）であり、確率・ベイズ事後分布・p値ではないことを明記

4. basis: supported (§4.5)

• basisはversion-relevant contentと明示 — hypothesis↔assumption変更にはversion bumpが必要

5. lag: unknown in feedback cycles (§6)

• unknown lagはlag.min==0として扱い（正の遅延を証明できないため）、全unknown辺サイクルはinstantaneous errorと明記
• 実装バグも修正済み（PR feat(causal): Phases 1–5 — check/analyze/diff, evidence, expectations, observation, ledger #369 commit f4bbc6a）

6. Evidence scope missing dimension (§4.4)

• 片方にしかない次元はunassessable、空scope（次元なし）は全体unassessableと明記
• missing scopeをuniversalとして黙認しないことを明示

commit c6fc9ea docs(causal): address PR #367 review

[rizumita]

再レビュー済みです。指摘した authority scope、claim version/retired lifecycle、missing scope dimension、lag:unknown の契約が #367 自身へ反映され、最新 main との競合も解消されています。CI 5件すべて成功を確認しました。

[rizumita]

最新 main (#411) の取り込み後も差分と競合解消を再確認しました。承認を更新します。