Fix Pi standard session mapping and shared evidence preservation - #67393

Pi's unified session parser emitted custom pi.* wrappers and duplicated streaming snapshots instead of standard conversation and tool observations. This PR maps recognized Pi output into the shared session vocabulary and closes diagnostic-attribution gaps found during the coordinated engine audits.

Approach

• Map Pi messages, reasoning, refusals, tools, progress, errors, and lifecycle observations to standard events. Reconcile partial/final snapshots without duplicating content or accounting; preserve exact text and supplied metadata.
• Preserve observed diagnostic status codes in unified payloads, declarations, and generated schemas. Keep child-agent errors out of the root execution aggregate without creating a universal parent-call alias or changing Pi tool nesting.
• Exclude quoted errors in observed message and tool output, result, and error payloads from raw diagnostic mining, including nested structured values, redaction, and partial framing. Explicit canonical root errors remain authoritative.
• Document run evidence and synthetic coverage boundaries. Historical artifacts are not rewritten, and unavailable observations do not produce fabricated tools, outcomes, or metrics.

The branch now includes the latest merged base. Common numeric-step, nullable-exit, rendering, and accounting prerequisites already landed through other engine PRs; they are not new changes in this PR's current diff.

Validation and review fixes

• Both Copilot findings were reproduced with five failing regressions, fixed in 10418f132d, replied to, and resolved. Coverage includes parent-call-only child diagnostics, root null/empty/invalid identities, all tool payload fields and nested values, genuine root-error controls, and Pi nested-tool scope.
• make fmt, full make lint, and make agent-report-progress passed. Existing lint warnings remain; no tooling or assertions were disabled. Declared Prettier 3.9.9 was selected instead of an ambient unpinned alpha.
• Typecheck, schema freshness, 1,621 session tests across 40 suites, 276 diagnostic/consumer tests, and the embedded-parser dependency check passed. The base merge also regenerated and compiled all 334 workflows.
• Four complete Pi traces replayed through canonical persistence and unified projection with exact content/accounting reconciliation and zero remaining pi.* events: historical success, provider failure, modern success, and failed workflow.

CI and evidence limits

The pre-fix merge-head CI snapshot passed all four CJS shards, typecheck, lint, and documentation build. Python CodeQL run 38045315822 failed during SARIF post-processing because the GitHub installation API rate limit was exhausted, after source analysis and SARIF export completed. That external failure is not locally reproducible. CI at the latest agent-pushed head is unverified and requires a maintainer re-trigger before merge; no workflow was dispatched or rerun.

Other engine fixes remain on their separate branches/PRs unless already merged into the base. Cursor stream-json coverage is documentation-derived and synthetic; Crush has no inspected successful executed conversation. Workflow failure is not proof of inference failure. Pydantic AI and generic-custom audits were archived before completion. AGY's initial canceled workflow-drift check was subsequently completed in its separate PR-finisher follow-up, including all 334 workflows; those supplemental AGY/shared-renderer commits are not incorporated into this branch. No PR was merged by this session.