[FEATURE] Cap the PR review input tokens (review_max_input_tokens) to stay under Groq's 8K TPM

[koydas]

🎯 Goal

Bound the size of the review LLM request so it always fits Groq's per-request TPM limit (8,000 on the free tier for openai/gpt-oss-120b, ADR-0025), the same way autofix_max_input_tokens already does for auto-fix (ADR-0017).

📍 Context

• Repo: autonomous-dev-loop
• Domain: LLM token budget
• Component: scripts/pr_review.mjs, config/models.yaml, scripts/lib/config.mjs
• Related: fix(groq): replace retired models with openai/gpt-oss-120b #161 (ADR-0025), feat(review): ground PR review verdict in executed tool evidence (ADR-0024) #160 (ADR-0024, tool evidence)

🚀 Description

Today the only cap on the review prompt is the diff's 12,000 chars (filterDiff). The PR body, dependency manifests, change classification, automation-gate context and the tool evidence from ADR-0024 are uncapped.

Measured on #161: input_tokens_est: 6743, plus review_max_tokens: 1024, totals ~7,767. That leaves about 230 tokens of margin under 8K. A PR with failing checks adds up to 2,000 chars of output tail per failing check (~500 tokens each), so it is the most likely case to get a 413 Request too large. After a 413 the review fails and no verdict is posted.

Proposal:

• New optional key review_max_input_tokens, validated by loadLLMConfig() like autofix_max_input_tokens. Default it so that system + input + review_max_tokens ≤ 8,000.

• pr_review.mjs assembles the user prompt within that budget, trimming in priority order:

  1. manifests
  2. PR body
  3. diff

  Tool evidence and the verdict instructions are never trimmed, since ADR-0024 relies on them.

• When anything is trimmed, the prompt states it, as the existing diffTruncated warning does.

• Log section token estimates in review.llm_request meta, like auto-fix's token_estimate.

🧩 Scope

• In: review stage budget, the config key, docs (docs/code-generation.md per-stage keys table, runbook 413 row), and a config test asserting the sum ≤ 8,000 using the prompt measured with estimateTokens().
• Out: generation stage input cap (separate issue if needed), tokenizer-accurate counting, Developer-plan tuning.

🧪 Acceptance criteria

• review_max_input_tokens is loaded and validated: absent, valid, and invalid (non-positive, non-integer) branches are tested.
• With a huge PR body, a large diff and 2 failing checks, the request built by pr_review.mjs stays ≤ review_max_input_tokens. Tool evidence is intact and the trim notice is present.
• The test fails if the system prompt + review_max_input_tokens + review_max_tokens > 8,000.
• Removing the key restores current behavior (no cap).
• Docs and CHANGELOG are updated. Amend ADR-0025 or add a new ADR.

⚙️ Constraints

• Must not drop or truncate the tool-evidence block or the verdict format instructions (ADR-0024).
• No new dependencies; chars/4 estimation (estimateTokens) is acceptable.
• Edit-guardrails: targeted edit in pr_review.mjs (≤ 30% of lines).

[github-actions]

✅ Issue Validation Report

Field	Value
Status	VALID — ready for development
Score	95/100 [██████████]

⚠️ Warnings (optional improvements)

• Edge case not covered: verify that when the combined token count is already within the limit, no trim notice is added and the request proceeds unchanged.

💡 Suggested Acceptance Criteria

Copy the block below and paste it into the issue description, then save. Saving will re-trigger validation automatically.

## Acceptance Criteria

- [ ] [ ] Given a PR with total token estimate below review_max_input_tokens, the generated request contains the full prompt without any truncation notice.
- [ ] [ ] When review_max_input_tokens is set to a non-integer value, loadLLMConfig() throws a validation error and the process aborts.
- [ ] [ ] If the tool evidence block exceeds its allocated space, the system logs a warning but does not truncate the block, preserving its content.
- [ ] [ ] The token estimation logged in review.llm_request matches the sum of system prompt, trimmed input sections, and review_max_tokens within a tolerance of ±5 tokens.