perf: cache DAG, batch top-k metadata, add full-pipeline recall benches

[JosephOIbrahim]

Summary

Three performance / measurement improvements from the second-pass MoE review (the "PR C" lane from the strategic split). Independent of PR #6 and PR #7 — no file overlap.

Fix	File	What it does	Why it matters
Cache cognitive DAG	src/mock_cogexec.py	Build the DAG + topological order once at module load; evaluate_dag reads from _DAG / _DAG_ORDER instead of rebuilding	The DAG shape is a constant (7 nodes, 5 edges). Sessions evaluate this on every exchange — ≥ hundreds per session — so N rebuilds become 1.
Batch top-k metadata fetch	python/harlo/encoder/__init__.py	Replace the per-id SELECT ... WHERE id = ? + fetchone() loop in semantic_recall with a single SELECT ... WHERE id IN (?,?,...); build {id: row} map and read in distance order	k+1 SQLite round-trips → 2 statements. Removes a measured N+1 (one SELECT for the full candidate scan, one per top-k).
Full-pipeline recall benches	crates/hippocampus/benches/recall.rs	Add three benches alongside the existing xor_search micro-bench: load_all_sdrs 100k, recall_full 10k k=5, recall_full 100k k=5	The existing bench measures only popcount + heap over pre-loaded candidates. The real recall path also does SQLite scan + 25 MB blob read. If load dominates, optimising popcount with SIMD is wasted work. Measure before optimising.

What's intentionally NOT in this PR

• F9–F12 from the second MoE review (anchor-buffer growth, S3 rupture-repair theatre, S8 sincerity gate, _CONSENT_SECRET regen). Each requires a constitutional-intent decision before any code change.
• Anything from PR fix(constitutional): S7 formula + Checkpoint TOCTOU + engine-lock timeout + HotStore commit handling #6 (S7 formula, Checkpoint TOCTOU, _engine_lock timeout, HotStore commit) or PR fix(salvage): apoptosis transaction + decompile listener-before-reset + audit regression test #7 (apoptosis transaction, decompile listener reorder).

Test plan

• cargo test -p hippocampus — 42 passed, 0 failed
• cargo bench --no-run -p hippocampus — new benches compile clean
• python3 -m py_compile on src/mock_cogexec.py and python/harlo/encoder/__init__.py — syntax clean
• Adjacent regression sweep (tests/test_motor/ test_brainstem/ test_elenchus/ test_hippocampus/ test_composition/ test_inquiry/ test_sync/ test_hot_store/) — 207 passed, 7 skipped (pre-existing env skips for pxr / mcp), 0 failed
• Cannot run encoder pytests in this sandbox — pre-existing env issue: python/harlo/encoder/semantic_encoder.py:12 imports numpy which isn't installed. The N+1 fix is syntax-clean and behaviourally equivalent (same ordering, same result keys, fewer round-trips); local CI with numpy can confirm.
• Want a real recall measurement. cargo bench -p hippocampus --bench recall -- --warm-up-time 1 --measurement-time 3 --sample-size 20 will produce numbers for load_all_sdrs 100k and recall_full 100k k=5 — those are the headline diagnostics this PR exists to enable.

Compliance

The 33 inviolable rules in CLAUDE.md are unchanged. R1 / R7 are pure perf; R8 is pure additive bench. Constitutional greps (sleep(, while True, DELETE.*audit, float32, cosine, reasoning_trace in elenchus/verifier.py) all return zero.

https://claude.ai/code/session_017arHKzx5mTUFiry7JhhRPs

---

Generated by Claude Code

Summary by CodeRabbit

• Tests

  • Expanded benchmarking suite with new performance measurement functions for recall pipeline operations and data loading scenarios.

• Refactor

  • Optimized database query performance through batch processing of trace metadata retrieval.
  • Improved computation efficiency via precomputed and cached execution graphs.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: d098aa0c-0c7d-429e-8c37-defe0b55dd0a

📥 Commits

Reviewing files that changed from the base of the PR and between 490e3e0 and b509cfa.

📒 Files selected for processing (3)

• crates/hippocampus/benches/recall.rs
• python/harlo/encoder/__init__.py
• src/mock_cogexec.py

---

📝 Walkthrough

Walkthrough

Three independent performance improvements: expanded end-to-end benchmarking for recall operations in Rust, batched SQLite metadata queries in semantic search, and precomputed DAG caching in cognitive execution evaluation.

Changes

Benchmark Suite Expansion

Layer / File(s)	Summary
Module Setup crates/hippocampus/benches/recall.rs	Module documentation and imports expanded to support both micro-benchmarks (xor_search) and full-pipeline benchmarks (load_all_sdrs, recall).
Helper Infrastructure crates/hippocampus/benches/recall.rs	Added build_db(n) function to construct and populate in-memory SQLite databases with TraceRecord entries.
Benchmark Functions crates/hippocampus/benches/recall.rs	Added bench_load_all_sdrs_100k, bench_recall_full_10k, and bench_recall_full_100k to measure full-pipeline performance across 10K and 100K trace databases.
Test Registration crates/hippocampus/benches/recall.rs	Expanded criterion_group! to register new load_all_sdrs and recall_full benchmarks alongside existing micro-benchmarks.

Trace Query Batching

Layer / File(s)	Summary
Batch Query Implementation python/harlo/encoder/__init__.py	Optimized semantic_recall metadata retrieval from individual SELECT ... WHERE id = ? queries to a single batch query using IN (...) clause, with dictionary-keyed assembly preserving distance-based ordering.

DAG Evaluation Caching

Layer / File(s)	Summary
Cache Infrastructure src/mock_cogexec.py	Added module-level private caches _DAG and _DAG_ORDER for precomputed graph and topological evaluation order.
Function Update src/mock_cogexec.py	Updated evaluate_dag() to reuse cached _DAG and _DAG_ORDER instead of rebuilding graph and topological sort on each invocation.
Import-Time Setup src/mock_cogexec.py	Added module initialization block to construct _DAG from build_dag() and compute _DAG_ORDER via topological sort at import time.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Possibly related PRs

• JosephOIbrahim/Harlo#3: Adds the initial xor_search micro-benchmark to the same recall.rs file; this PR expands that benchmark suite with full-pipeline measurements and supporting infrastructure.

Poem

🐰 Three swift hops through the code we made,
One batches queries, one caches cascades,
And benchmarks now measure from start to the end—
Performance and wisdom, bundled as friends! ✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 30.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately summarizes all three primary changes: DAG caching, batch metadata fetching, and new benchmark functions for the full recall pipeline.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch claude/perf-measurement

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}