Separate validation evidence gaps from correctness failures

[yyjeqhc]

Summary

• separate structured validation target identity from request-scoped require_tests / min_tests evidence assertions
• preserve immutable raw ToolCall history while projecting successful validator execution plus insufficient count proof as an evidence gap rather than a correctness failure
• reconcile current validation actionability, historical failures, and unexpected tool failures through the same canonical projection used by handoff and finish_coding_task
• preserve mutation-start boundaries, exact target/project reconciliation, expected/observe neutrality, and outcome_unknown fail-closed behavior
• align closeout schemas, tests, and maintainer/user documentation with the new semantics

Review follow-ups

The branch was independently reviewed after rebasing onto current main. The review fixed two duplicate expected_generation test fixtures introduced by overlap with #316, renamed a stale test that still described count insufficiency as failed Session validation, and corrected stale architecture/schema documentation. No larger redesign blocker was found.

Validation

• cargo test -p webcodex-validation — 92 passed
• cargo test -p webcodex-workflow-session — 165 passed
• cargo test -p webcodex finish_coding_task_ — 25 passed
• cargo test -p webcodex failure_history_ — 8 passed
• cargo test -p webcodex validation_handoff::cargo_test_assertions:: — 4 passed
• cargo test -p webcodex-tool-contracts — 88 passed; focused post-review schema checks 2 passed
• cargo test -p webcodex-tool-runtime-contracts — 68 passed
• cargo test -p webcodex-core validation_identity — 1 passed
• cargo check --workspace --all-targets — passed
• cargo check -p webcodex-tool-runtime-contracts --all-targets — passed after reviewer fixture cleanup
• cargo fmt --all -- --check — passed
• git diff --check origin/main...HEAD — passed
• workspace hygiene — clean; active Jobs — 0

Closeout dogfood keeps historical failure evidence but reports it non-actionable (actionable_unexpected_count=0, no unresolved validation failure) instead of hard-blocking the current clean workspace.