feat(e2e): support forbidden recall context

[russell001209-ai]

Which issue or RFC does this PR close?

Refs #1263.

Follow-up to #1272.

Rationale for this change

PR #1272 added declarative recall probes but deferred negative benchmark cases because the evaluator could not express context that must not be returned. Without a negative contract, a probe can satisfy its required fragments while still returning stale or otherwise forbidden context.

What changes are included in this PR?

• Add an optional forbidden_context contract to recall probes.
• Match required and forbidden fragments case-insensitively with Unicode normalization.
• Reject blank fragments and contracts where a forbidden fragment is contained in required context.
• Treat any forbidden match as a hard acceptance failure, independent of probe_coverage.
• Record the pre-redaction Boolean verdict for replay scoring and discard matched prepared context before persistence.
• Preserve compatibility with existing manifests and older replays that do not contain the new field.
• Document the contract in the Bub README and the English and Chinese RFCs.

The related production substring paths were checked: recall-probe matching is updated here, while expected_memory remains a separate Memory-recording contract and is unchanged.

Are there any user-facing changes?

E2E workload authors can now declare context that a recall probe must not return. Existing manifests keep their previous behavior because forbidden_context defaults to empty. Existing evaluation reports are unchanged unless a workload declares a forbidden fragment. There are no PowerContext Server, Client, or persisted Memory API changes, and no new model calls or scheduled work.

How was this change tested?

• uv run --locked ruff check e2e/bub
• uv run --locked ruff format --check e2e/bub
• uv run --locked ty check --project e2e/bub --python e2e/bub/.venv --python-version 3.12 e2e/bub/src integrations/bub/src
• uv run --project e2e/bub python -m pytest e2e/bub/tests -q (23 passed)
• uv run --project e2e/bub powercontext-e2e --help
• git diff --check

Regression coverage includes a forbidden match with probe_coverage: 0, canonically equivalent Unicode, manifest validation, removal of matched context before replay persistence, and live/offline replay parity for both recorded true and false verdicts.

AI usage statement

OpenAI Codex with GPT-5.6-sol was used for implementation assistance and independent adversarial review. The final diff passed the repository checks listed above, and the final independent review reported no remaining findings.

[PsiACE]

lgtm