General Review

[EnRaiha]

Context

AI Review Pipeline — 10-part NotebookLM snapshot via repomix + Gemini 3.8 Flash (extended thinking), reviewed against origin/main @ cc97b0885 (latest: 7205d9fd9).

Verified bugs (9) — each reproduced with a failing test on origin/main

ID	Bug	Evidence	Severity
D1	Governor pressure depth compounds: 6× Critical → depth 1 (should settle 32)	core_loop/pressure.rs:123 spsc_critical_ticks_do_not_collapse_to_one	sev:3-medium
D2	MemoryGovernor::try_reserve UnknownEngine leaks global counter	nodedb-mem/src/governor.rs early-return after global increment — try_reserve_unknown_engine_rolls_back_global leaks 1000 bytes	sev:2-high
D6	KV secondary index rewrites on identical update	engine/kv/index/set.rs:144 unconditional remove+insert — d6_identical_update_elides_index_writes → 2 writes (expect 0)	sev:3-medium
S1	LIMIT -1 silently unbounded	nodedb-sql/src/coerce.rs Option-based helper → None — s1_negative_limit_becomes_unbounded	sev:3-medium
S2	ANALYZE users(id) absorbs (id) into collection name	ddl_ast/parse/maintenance.rs:14 — s2_analyze_parenthesized_payload_absorbs_token	sev:4-low
V3	Cost planner oversample 3 vs BBQ default 4	planner/cost.rs vs rerank/codecs/bbq.rs:88 — v3_oversample_constant_matches_bbq ratio 3 ≠ 4	sev:4-low
C1	TOML ${VAR} never expanded (literal /\${VAR})	config/server/config.rs from_file — c1_unresolved_placeholder_fails_startup	sev:3-medium
C2	Env cluster overlooked when [cluster] absent (NODEDB_NODE_ID dropped)	config/server/env/cluster.rs — c2_env_cluster_auto_initializes_section	sev:3-medium
C3	Malformed numeric env warns + falls back	config/server/env/numeric.rs — c3_malformed_env_errors_instead_of_silent_default	sev:4-low

PRs:

• PR fix(sql,vector): reject negative LIMIT, bound ANALYZE name, unify oversample #270: S1, S2, V3.
• PR upcoming config: C1.
• PR upcoming memory/dataplane: D1, D2, D6 (C2/C3 deferred per decision).

Recommendation

• Kernel tiers (suggestion, not bugs): fused BBQ dequantize-distance with HAL — (a) avx512 512-bit, (b) AVX10 256-bit, (c) avx2 / neon / wasm simd128 — see DIGEST-PDF-20260904.md V1 2a/2b/2c.

PRs reference this issue.

[farhan-syah]

What is this?

[EnRaiha]

What is this?

A mess of early push where contributer too eager and ai assistant not ready with pr

[farhan-syah]

Please close this and refile it as separate issues.

Nine unrelated findings across nodedb-mem, nodedb-sql, nodedb-vector, config, and the KV engine sit in one issue. No single change closes it, no one reviewer owns all of it, and the tracker cannot show which parts are fixed. Three PRs already point here, which means this issue will stay open long after its contents are resolved.

One finding, one issue

File each through the bug report template. It asks for the version, the repro, expected vs actual, and severity — fill every required field.

Write each issue for someone who has never read this thread. They must be able to reproduce the defect from the issue body alone, with no other document open.

Titles

The title names what is broken and what it costs. It does not name the process that found it.

Instead of	Write
AI Review Pipeline: ... review findings	MemoryGovernor::try_reserve leaks the global counter when the engine is unknown
D2 / S1 / C1	LIMIT -1 is silently treated as unbounded instead of rejected
Config bug	ANALYZE users(id) parses the parenthesised payload into the collection name

Internal shorthand — D1, S2, V3 — means nothing to a reader. Drop it.

Keep every issue self-contained

• Never cite a file that is not in the repository. DIGEST-PDF-20260904.md is referenced in this issue and in fix(sql,vector): reject negative LIMIT, bound ANALYZE name, unify oversample #270. No reader can open it, and the reference will still be dead a year from now. Quote the relevant finding into the issue body instead.
• No internal paths, working documents, snapshots, or tooling names. The issue tracker is public. Anything a reader cannot follow does not belong in it.
• Pin one commit. This issue names both cc97b0885 and 7205d9fd9. Pick the one you tested and put it in the version field.

Evidence is a repro, not a line number

core_loop/pressure.rs:123 is already wrong — the file is nodedb/src/data/executor/core_loop/pressure.rs, and any line number drifts within days. Give the repro instead:

• SQL-reachable defects (LIMIT -1, ANALYZE users(id)): copy-pasteable SQL from a fresh data directory, annotated with what each statement returned.
• Internal defects (try_reserve rollback): the smallest sequence of calls that shows it, plus the observed and expected state.

A file path is context. It is never the evidence.

Severity belongs to each issue, not to the batch

Set it per issue against the template's definitions, and let triage adjust it.

Sort the list first — three of these are not bugs

Finding	File as
try_reserve leaks the global counter on unknown engine	Bug. Highest impact of the nine — every other error path in that function rolls back; this one returns after three counters are incremented. File this one first.
LIMIT -1 unbounded	Bug.
ANALYZE users(id) name absorption	Bug. Cosmetic severity.
Governor pressure depth compounding	Bug, once it carries a repro.
Cost planner oversample constant vs BBQ default	Bug if the two are required to agree. State why they must, or it is a chore.
Env cluster section, malformed numeric env	Bug each, with a repro showing the dropped setting.
TOML ${VAR} expansion	Feature request. Environment interpolation is not implemented and nothing documents it, so nothing is broken.
KV index rewrite on identical update	Feature request or type:chore. The writes are correct; eliding them is an optimization. Filing it as a bug with "expected 0 writes" asserts an invariant that does not exist.
Fused BBQ dequantize-distance kernel tiers	Feature request. It is a proposal, not a finding, and it does not belong in a defect list at all.

PR #270

Keep it — the three fixes are sound and the tests are the right shape. Two changes to its description:

1. Remove the DIGEST-PDF-20260904.md reference and the pipeline framing. State the defect and the fix.
2. Link the three new issues with Fixes #<n>, one per finding, so each closes on merge.

Do the same for the two upcoming PRs: one issue per fix, linked by Fixes.

[EnRaiha]

Closing per maintainer review (farhan-syah) — refiled one finding per issue, template-complete:

Bugs:

• MemoryGovernor::try_reserve leaks the global counter when the engine is unknown #271 MemoryGovernor::try_reserve leaks the global counter on unknown engine
• LIMIT -1 is silently treated as unbounded instead of rejected #272 LIMIT -1 silently unbounded
• ANALYZE users(id) parses the parenthesised payload into the collection name #273 ANALYZE users(id) name absorption
• SPSC read depth compounds to 1 under sustained Critical memory pressure #274 SPSC read depth compounds to 1 under sustained Critical pressure
• Vector cost planner assumes oversample 3 while the BBQ quantizer trains with 4 #275 Vector cost planner oversample vs BBQ default
• NODEDB_NODE_ID and NODEDB_SEED_NODES are dropped when the config has no [cluster] table #276 NODEDB_NODE_ID dropped without [cluster] table
• Malformed NODEDB_DATA_PLANE_CORES falls back to defaults instead of failing startup #277 Malformed numeric env falls back instead of failing startup

Features/chores:

• Support expansion in TOML config files #278 ${VAR} expansion in TOML config (feature)
• Elide redundant secondary-index rewrites on identical in-place updates #279 Elide identical KV index rewrites (chore)
• Fused BBQ dequantize-distance kernels with tiered SIMD dispatch #280 Fused BBQ SIMD kernels (feature)

PR #270 carries the three fixes (S1/S2/V3) with Fixes links.