fix: allow clean-run summaries to bypass never_worse guard for injected JSON (fixes #4218)

[TechWizard9999]

Summary

Part of #4218. This PR fixes the case where rtk ruff check on a clean tree prints the raw injected JSON [] instead of the filter's human summary "Ruff: No issues found".

Root Cause

When rtk injects --output-format=json (or similar), the tool's empty-run output is extremely short (e.g., [] = 2 bytes). The never_worse guard compares token estimates against this injected output and rejects the filter's informative summary because ceil(21/4)=6 > ceil(2/4)=1.

Changes

1. src/core/runner.rs: Added empty_json_patterns field to RunOptions and logic to prefer the filter's summary when raw stdout matches a registered empty-run pattern.

2. src/cmds/python/ruff_cmd.rs: Registered [] as ruff's clean-run empty JSON pattern.

3. src/cmds/php/pint_cmd.rs: Registered {"files":[]} as pint's clean-run empty JSON pattern. Dropped in the maintainer fixup: pint runs under a tee label, and its clean output already loses to pint: ok.

Testing

• The existing filter unit tests pass (they verify the filter produces the correct summary)
• The fix ensures the guard doesn't override the summary when raw output matches the registered pattern
• Other format-injecting tools (rubocop, rspec, sqlfluff, phpstan, golangci-lint, go test, gh/glab) can be updated similarly in follow-ups

[rtk-wshm-sync-bot]

wshm · Automated triage by AI

📊 Automated PR Analysis

🐛 Type	bug-fix
🔴 Risk	high

Summary

Fixes issue #4218 where rtk's never_worse guard rejected the filter's human-readable summary in favor of raw injected JSON output (e.g. '[]') because the raw output was shorter, by letting ruff/pint register known clean-run empty-JSON patterns that bypass the guard. The diff also includes an unrelated change to src/core/retriever.rs that reverses a previously documented invariant (tee mode no longer avoids creating a sqlite artifact), which is not mentioned anywhere in the PR description.

Review Checklist

• Tests present
• Breaking change
• Docs updated

Linked issues: #4218

---

Analyzed automatically by wshm · This is an automated analysis, not a human review.

[TechWizard9999]

@KuSh —Could you please review this

[KuSh]

Round 1 of 3 — APPROVED

Claim: when rtk injects --output-format=json into rtk ruff check and the tree is clean, show the filter's Ruff: No issues found instead of ruff's raw [], which never_worse picked because it is shorter.
Scope: accept if narrowed to "the #4218 fix for rtk ruff check" (round 1, frozen). The two #4219 commits (9833d664, 637a2d18: the tee path creating the recall store) are the whole of #4263 and are reviewed there; #4219 is an open design question, since today a tee install leaves no sqlite artifact behind. The pint registration is dropped: pint never had the symptom.
Ran: rtk ruff (ruff 0.16.6) × 11 invocation forms (clean, dirty, user --output-format=json in both spellings, --output-format=gitlab, --exit-zero, --fix --exit-non-zero-on-fix, missing path, config error, bare check, format --check --output-format=json) × develop c75f1590 / head 7a446f14 / fixup bff334ca; rtk pint --test on pint 1.32.1's real clean JSON, captured in Docker and replayed; rtk grep in tee mode for recall.db creation. Invariants checked: under the hook (ruff check and uv run ruff check rewrite to rtk ruff check), every branch of the new decision (injected clean, user JSON, other formats, non-check subcommands, tee label), split streams (config error on stderr), and the accounting row. The clean row now books 1 → 6 tokens (−5). That is the intended trade: tracking keeps negative savings as they are. Every other row is unchanged against develop. An independent run of the fixup covered about 35 more forms (every --output-format value, --statistics, --diff, --add-noqa, -o, format --diff, stdin, stderr warnings): the fixup differs from develop only where stdout is exactly an injected [].
Matrix: executed this round
CI @bff334ca: all green (11/11; check-target skipped)

Blocks merge (2, frozen at round 1): both closed by the fixup

1. src/cmds/python/ruff_cmd.rs:114: the bypass fires for JSON the user asked for. rtk ruff check --output-format=json clean.py | jq length prints 0 on develop and fails with jq: parse error on the head, because [] is swapped for the summary even when the user chose the format. Fixed by registering [] only when is_check && !user_set_output_format, i.e. only when rtk injected the format. Both spellings are back to [] on the fixup.
2. src/core/retriever.rs:387,433: out-of-scope behaviour change. In tee mode, rtk grep now creates recall.db (develop: no sqlite file). This reverses the invariant documented in the comment it replaces, and it is exactly #4263's change, so the fixup reverts it here and #4219 is decided on #4263.

Optional: will not hold merge

• The pint registration {"files":[]} was never read: pint runs with .tee("pint"), and the tee branch goes through print_with_hint, which never looks at the new field. Pint also never had the symptom: a clean pint 1.32.1 run prints a ~140-byte JSON (about, files, time, memory) that already loses to pint: ok. Dropped in the fixup, and the field doc now says it is not consulted under a tee label.

• The decision moved into guard_stdout and got two tests. A mutation in each direction fails the matching test. guard.rs now lists this exemption next to the rtk diff one.

• After --fix has fixed everything, the summary reads Ruff: No issues found. Develop printed [], and ruff's own "N fixed" line is lost on both sides in JSON mode. The exit code is propagated. A filter wording such as "no remaining issues" under --fix would be more exact; that is a nit for a follow-up, not this PR.

• Every clean rtk ruff check now books a negative savings row (1 → 6 tokens). That is the intended trade: tracking keeps negative rows as they are.

• The option is renamed clean_outputs to match what it holds (exact outputs, not patterns).

Filed as follow-ups

• #4218 stays open (the body now reads "Part of #4218"): rtk lint ruff on a clean tree still prints [] (it goes through emit_guarded, not the runner), and the other injectors the issue lists are unchecked. Evidence added to #4218 (comment).

Checked and correct — no need to re-verify

• Dirty tree, missing path, --exit-zero, bare check: same output and exit code as develop.
• A config error (stdout empty, error on stderr) is forwarded unchanged.
• --output-format=gitlab and ruff format --check --output-format=json print [] on a clean tree and are unchanged: they are not registered, and their filters echo it anyway.
• CRLF: []\r\n trims to the pattern (pinned by a test).

Your questions

• (none; the review request is this round)

Next

fixup f9f40924 + bff334ca pushed: reverts the #4219 commits (kept on #4263), limits the bypass to the format rtk injected, drops the pint no-op and adds tests. Approving on green CI.
Rounds: 1/3. Threads: 0 open, 0 resolved this round.