Add bounded parallel reconstruction, Grok Build, and streaming agent monitoring

[SamG-Coder]

Reconstruction runs can now overlap independent functions with isolated provider state and durable recovery, while the existing monitor streams progress and exposes agent code and activity. Sequential execution remains the default, and candidate acceptance checks remain explicit.

Depends on #15. Please merge #15 first; this branch includes its prerequisite performance and monitor changes.

Changes above #15:

• Run local candidate validation before model review, reject unresolved executable decompiler output, provide assembly evidence, preserve compiler diagnostics, and stop repeated repairs when evidence scopes conflict.
• Add the Grok Build CLI provider with native conversation resume, prompt files, usage/error metadata, isolated configuration, existing authentication-path reuse, and bounded process cleanup. Tool-free reconstruction requests retain the one-turn limit and disable native subagents.
• Support 1–32 function workers with separate model-request and validation limits. Per-function providers, conversations, budgets, artifacts, and caches remain isolated; shared backend operations are serialized.
• Add cancellable FIFO request admission, optional budgeted retries, HTTP 429 cooldown, and separate queue/request timing. Disable hidden SDK retries.
• Coordinate dependency-aware scheduling, session leases, durable checkpoints, interrupted-budget recovery, and deterministic cumulative proposal validation and promotion. Failed or cancelled proposals cannot publish provisional successes.
• Extend the existing monitor to consume cooperative batch progress and manage configured workers, including resume, stop-file clearing, ownership checks, and duplicate-start prevention.
• Deliver an initial snapshot followed by WebSocket updates, with reconnection and source freshness reporting. Add bounded native JSONL ingestion, agent and batch selection, historical views, and terminal-style Code / Activity / Full response tabs with scroll following and keyboard navigation.
• Keep unattributed tokens in a shared stream rather than assigning them to a child. Display compiled drafts separately from reviewed or behaviorally verified results.
• Allow explicitly allocated short Grok homes for native callers, and check a conservative Windows session-path budget before launch. Preserve managed requirements without copying credentials or overwriting existing homes.

Validation at a88db86:

• Full local suite: 293 passed, 9 skipped.
• Ruff passed for src and tests; Mypy passed with Linux platform stubs across 80 source files.
• Coverage includes parallel scheduling, dependency cycles, leases, budgets/retries, process cancellation, recovery/promotion, monitor controls, real WebSocket snapshot/update delivery, origin checks, bounded historical access, partial event lines, agent attribution, and short-home/path-budget behavior.
• Focused JavaScript checks covered terminal following, scrollback pause/resume, tab selection, and keyboard navigation.
• A real two-subagent persistence probe completed successfully with both metadata and non-empty output files saved, exit code 0, and empty stderr after shortening the session home. This verifies persistence, not reconstruction equivalence.

Scope and limitations:

The native batch runner used for local experiments and its project-specific recovery logic are outside this PR. This PR adds reusable provider/monitor support; it does not introduce native batch orchestration as a supported ReAgent reconstruction mode. Unattributed Grok token streams may interleave; per-child responses are shown only when identified by provider task IDs. Compilation and model review do not establish binary equivalence.

No project-specific reconstruction files, run reports, credentials, or live-test transcripts are included. See docs/parallel-functions.md, docs/grok-build.md, and docs/live-monitor.md for configuration and limitations.