neohaskell/docs/archive/2026-07-ai-artifacts/claude-skills/integration-pipeline-preview/13-performance-impl/02-deep-audit

ARCHIVAL KNOWLEDGE — DO NOT USE OR REFERENCE. Superseded by codemap/ + root AGENTS.md. Scheduled for deletion once the pipeline is verified (Phase 6). Manifest: docs/archive/2026-07-ai-artifacts/MANIFEST.md

name: 02-deep-audit description: Reads the implementation source and applies the performance methodology to produce file-line findings. kind: leaf executor: sonnet model: claude-sonnet-4-6

Deep Performance Audit

Reads changed source files plus siblings and applies every section of the performance methodology, producing raw findings with file:line locations.

Review posture

Assume the artefact under review was produced by a language model (ChatGPT-class output). Treat plausible-looking claims as unverified, expect hallucinated APIs and missed constraints, and refuse to pass anything not directly traceable to the methodology in ../../references/performance-methodology.md and the actual source lines. Strict review is the default — benefit of the doubt goes to the rubric and to the source, never to the producer.

Inputs

• stdin — JSON array of static-scan findings from step 1.
• Changed source files under module_path and their siblings.
• ../../references/performance-methodology.md
• ../../references/nhcore-context.md

Plan

1. Read stdin → verify: array parses.
2. Read changed files and siblings → verify: at least one file loaded.
3. Read the performance methodology and nhcore context → verify: both exist.
4. For each methodology section, ask its question against the code → verify: every section produces at least a "not applicable" or real finding with file:line.

Assumptions:

• Static-scan findings are folded in, not replaced.
• Every finding includes file:line; lines are 1-indexed.
• The 50k req/s target frames severity.

If any assumption fails, refuse — do not guess.

Steps

1. Parse stdin into the static-scan finding list.
2. Resolve and load the changed files plus immediate siblings.
3. Load the performance methodology and nhcore context.
4. For each methodology section (1-9), walk the code and produce findings.
5. For each finding, populate severity (Blocking/Advisory/Informational), rule (e.g. inline-pragmas, tojson-without-toencoding, tvar-map-contention), location (file:line), message, recommendation.
6. Normalise every finding's location first: if a static-scan entry carries file and line instead of location, set location = "<file>:<line>" before dedup. After normalisation every entry has the same location shape (file:line).
7. Merge static-scan findings with deep findings (deduplicate by the normalised rule + location pair).
8. Emit the merged JSON array on stdout.

Output

JSON array of merged static + deep performance findings on stdout.

Refusals

• Methodology or context reference missing → refuse: "performance methodology or nhcore context missing".
• No changed files resolved → refuse: "no changed Haskell files".