feat: simplify Context Tree value audit evidence model

[Gandy2025]

Summary

• simplify the semantic model to three user-facing concepts: Task, Read, and optional Effect
• replace the v0.2 exposure/confidence/support/sampling model with schema-v2 Task judgments
• keep the collector and read-only authorization boundary unchanged
• report every available Task without a minimum quota, task-type gate, expansion batch, or saturation state
• preserve historical report reproducibility when the bound Tree advances after collection
• bump the Skill payload to 0.3.0 and update its docs, eval checklist, and tests

Effect rule

An Effect is recorded only when a relevant normal-content Read precedes the choice, the later choice or outcome reasonably shows confirmed, constrained, redirected, or conflicted influence, and no more direct instruction or evidence fully explains the result. A decision receipt remains supporting evidence, not causal proof.

Validation

• repository Skill validator: passed
• upstream skill-creator quick validator: passed
• Python compile check: passed
• full test suite: 47 passed
• git diff --check: passed
• fixed seven-day historical rerun: 44 clear Tasks; 26 observed Reads; 18 unresolved Reads; 16 Effects; 10 observed Reads without an Effect
• Effect distribution: constrained 8, redirected 5, confirmed 2, conflicted 1
• two independent review passes completed: model/compatibility and report semantics/safety

QA case disposition: no change. The deterministic behavior is covered by the existing repository suite; the private historical rerun remains an acceptance artifact rather than a committed fixture.

Context Tree

Constrained by agent-team-foundation/first-tree-context@0c7201e35de8abfcf1f755a4e9f7f4c1c439d9e5, especially system/context-management/task-skill-design.md: decision receipts are Agent attribution and must not be treated as server-verified causality.

[yuezengwu]

Reviewed head 9f426038e0b3204edfdc9e47c636d65b478748cc.

The change consistently replaces the v0.2 exposure/confidence/support/sampling model with schema-v2 Task, Read, and optional Effect judgments while leaving collection and authorization unchanged. I checked the reporter's Task-window/reference validation, Read/choice ownership, single-Effect shape, legacy-field rejection, aggregate conservation, and collection-time Tree snapshot handling. The bound Context Tree contract that a decision receipt is attribution/supporting evidence—not causal proof—is preserved in the Skill and report rules.

No blocking issue found. Non-blocking wording follow-up: README's pipeline section still says unsupported Runtime history “stays pending”; the new formal model calls the affected Read unresolved, so aligning that sentence would avoid carrying an old-state term forward.

Diff-only review; I did not rerun tests or QA.

[Gandy2025]

Addressed in c18a674c4ff9bc5777858bc7e5b6c75844c6238e.

• Replaced the README's unsupported-Runtime “pending” wording with the formal v0.3 result: affected Reads are unresolved.
• Aligned the same wording in the Runtime support matrix and manual behavior checklist so the old state term does not survive elsewhere.

Validation:

• repository Skill validator
• upstream skill-creator quick validator
• repository contract tests: 2/2
• git diff --check

[yuezengwu]

Re-reviewed head c18a674c4ff9bc5777858bc7e5b6c75844c6238e after the documentation follow-up.

The incremental commit consistently replaces the remaining unsupported-Runtime “pending” wording with the schema-v2 Read result unresolved in README, the Runtime adapter matrix, and the manual checklist. It resolves the prior advisory without changing runtime behavior or introducing a new issue.

No blocking issue found. Diff-only follow-up review; I did not rerun tests or QA.