feat(validation): add repeatable runs with process telemetry

[morluto]

Description

A single validation run cannot distinguish a stable result from a flaky one, and the existing runner records only exit/output timing. It cannot answer whether repeated attempts agreed, where a timeout occurred, how much process-tree resource was used, or whether descendants survived shutdown.

This adds bounded validation run groups with explicit attempt count, concurrency, per-run and overall timeouts, output limits, sampler interval, and MCP readiness timeout. Every attempt remains an ordinary persisted validation run; the group adds aggregate classification, pass/fail/error/cancel counts, timing distribution, peak resource metrics, and ordered attempt references.

Process telemetry uses maintained gopsutil APIs and binds samples to PID plus creation time to avoid PID-reuse errors. Unsupported metrics are represented as unavailable with a reason, never fabricated as zero. Process-group cancellation and a bounded shutdown check report surviving descendants separately from semantic test classification.

Declared mcp_stdio validations use the official MCP Go SDK for initialize and tools/list milestones. Protocol readiness is not inferred from arbitrary stdout, and readiness/shutdown timeouts abort the owned transport without an unbounded wait.

CLI and MCP expose explicit repeat operations; defaults are applied at runtime so omitted schema fields behave consistently across clients.

Closes #115.

Suggested review order

1. internal/evidence/repeat.go for bounded orchestration and aggregation.
2. internal/evidence/telemetry.go and runner changes for process identity, metrics, and cleanup.
3. internal/evidence/mcp_runner.go for SDK-backed protocol milestones and shutdown.
4. Persistence, application/CLI/MCP adapters, tests, and documentation.

Testing

• make verify
• make test-race
• Windows cross-compilation for evidence, application, CLI, and MCP packages
• Regression coverage for repeat ordering, cancellation, timeout phases, partial metric availability, descendant cleanup, runtime defaults, and repeated MCP readiness/shutdown behavior

Compatibility and scope

• Storage and side-effect invariants remain intact
• No unrelated cleanup or generated-output churn is included

Existing single-run behavior remains available. Existing corpora require the explicit backed-up migration flow. This branch, #118, and #122 currently occupy the next Goose migration slot; whichever merges later must rebase and renumber its migration before merge.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.