feat(one-agent): make one job = one one call a first-class RAS rule

[yaonyan]

Summary

Addresses the gap reported in #8 (comment): the resident prompt told models to batch but the model still opened multiple one calls per job, losing the control node where reason() belongs. The old wording ("One bounded action phase = one RAS call") was buried in Directives and abstract; this PR rewrites the mental model using the round-trip cost that models already understand from ReAct training data, and backs it with a stricter control-node eval.

Changes

• src/prompts.ts — rewrite the resident mental model:
  • Frame one as a working session (sandbox), contrasted with ReAct's reason-act-observe round-trip tax; every one call appends raw output to the shared conversation and consumes attention budget.
  • "One job = one one call" is a first-class planning default: batch all currently known related work into one bounded call. Starting a new call is allowed only when the previous one failed, timed out, or returned evidence that reveals a genuinely new target or dependency.
  • reason() is judgment-only at genuinely uncertain convergence points (targeting, branching, retry-vs-escalate, classification, synthesis); deterministic policies (thresholds, exit codes, rule-based transforms) stay in code/shell.
  • Inputs contract aligned across bash/python/typescript mode deltas: only values piped into act <tool> - must be JSON objects matching the tool schema; strings, arrays, and primitives are allowed elsewhere.
• one-runner-bash.md / one-runner-python.md — same one-job-one-call rule and inputs contract for on-demand runner docs.
• src/ras/*.ts — inputs schema descriptions match the relaxed contract for bash, python, and typescript.
• evals/agent_eval.ts:
  • Evidence-stream counting now strips comments and quoted string literals first, so act bash '{"command":"git log"}' counts one stream (not two) and cat/git mentions in prompts no longer inflate the batch score.
  • converge requires the exact normalized file set plus a non-empty reason (no more substring superset pass).
• evals/datasets/ras-control-node.json + evals/fixtures/changed-files.txt — add REPL-style cn-repl-converge probe forcing 2–3 evidence streams to converge in one call (expectBatched: true).
• tests/prompts.test.ts — assert the new mental-model and inputs-contract wording across all three modes.
• vitest.config.ts — retry is CI-only so local flaky regressions surface immediately.

Validation

• pnpm --filter=@one-agent/agent test → 108/108 pass (Test step runs in CI).
• tsc --noEmit clean.
• Eval run (deepseek-v4-flash):
  • ras-control-node: model now batches the initial evidence read (one call reads all 3 fixtures) vs. the previously observed 5 separate round trips; the strict eval still flags extra follow-up calls, which is the regression signal we want to keep.
  • directanswer gate unchanged vs. baseline (no regression).

[yaonyan]

Review fixes (commit 41e8686)

Addressed all Bugbot findings:

• [P1] Evidence-stream double counting — countEvidenceStreams now strips comments and quoted string literals before counting (stripCodeNoise), so act bash '{"command":"git log"}' counts one stream instead of two, and cat/git mentions in prompts or comments no longer inflate the batch score.
• [P1] converge false positives — now requires the exact normalized file set (got.size === want.size + full match, no substring superset) and a non-empty reason, so returning all files or not-src/b.test.ts.bak no longer passes.
• [P2] "One job = one call" as an absolute constraint — reworded to a planning default with explicit recovery exceptions: start a new one call only when the previous one failed, timed out, or returned evidence revealing a genuinely new target or dependency. Synced across prompts.ts and both runner docs.
• [P2] inputs values must be objects — narrowed to "only values piped into act <tool> - must be JSON objects matching the tool's argument schema; strings, arrays, and primitives are allowed elsewhere". Applied consistently to the bash/python/typescript mode deltas, the src/ras/*.ts schema descriptions, and one-runner-bash.md.
• Non-blocking: global retry: 1 — now process.env.CI ? 1 : 0 so local runs surface flaky regressions immediately.

Validation: pnpm --filter=@one-agent/agent test 108/108 pass, tsc --noEmit clean, CI green on this head (build job includes build/typecheck/Test).