feat: sandcastle refinement loop with critic-based convergence

[jerome-benoit]

Description

Replace single-pass implement→review→merge with a modular iterative implement↔critic refinement loop. Each task gets its own parallel sandbox with convergence detection, quality ratchet, and automated PR creation.

Architecture

Planner (opus) → selects issues
  For each issue (parallel, max 3):
    Sandbox with implement↔critic loop:
      Implementer (sonnet) → codes + commits + pushes
      Critic (sonnet) → structured findings JSON (nonce-tagged)
      Dedup (context-hash) → convergence check
      Quality ratchet → rollback on regression
      Best-state checkpoint → restore optimal intermediate
      Validation-in-loop (ARCS) → deterministic convergence
    Finalize: validate → rebase → PR (draft if non-converged)

Key Design Decisions

• Flat iteration budget (50/round) — evidence: ARCS, SWE-Agent, AutoCodeRover all use flat
• Context-hash dedup (±3 lines SHA-256) — drift-safe, CodeQL/Qodana pattern
• Severity-weighted convergence — refuses convergence if CRITICAL/HIGH persist (OpenHands)
• Best-state tracking — resets to best intermediate on non-convergence (SWE-Agent)
• Validation-in-loop — deterministic convergence when tests pass (ARCS)
• Async subprocess execution — util.promisify(execFile) unblocks event loop for true parallelism
• Nonce-tagged critic output — prevents injection from code content
• One PR per task — no batch merge, each issue gets its own PR

Modules

File	Lines	Responsibility
constants.ts	64	Shared constants + execFileAsync + getHeadSha + toErrorMessage
types.ts	83	Zod schemas + exported interfaces + parseFindingsSafe
concurrency-pool.ts	69	O(1) FIFO semaphore (linked list)
task-source.ts	248	TaskSource interface + GithubIssueSource (fetch + sanitize + plan)
refinement-loop.ts	580	Core loop: implement↔critic + dedup + ratchet + convergence
finalizer.ts	281	Validate + retry + rebase + push + PR creation
main.ts	103	Thin orchestrator: discover → pool → loop → finalize

Prompts

Prompt	Role	Key rules
plan-prompt.md	Issue selection	Prefer single-file scope, exclude blocked
implement-prompt.md	Code + commit + push	Cross-validate findings, full validation before push
critic-prompt.md	Structured review	≤5 HIGH/CRIT findings, nonce-tagged JSON, known decisions blocklist

Type of Change

• New feature (non-breaking change that adds functionality)
• Refactoring (no functional changes)

Checklist

• I have run npm run type-check && npm run test && npm run prettier-check && npm run lint
• I have run npm run build && npm run check-build && npm run build:v2 && npm run check-build:v2
• My changes follow the existing code style
• E2E tested locally (planner + parallel implementers started successfully)

Related Issues

Fixes #110