Local stress r1 correction ledger¶
The dedicated fix/stress-benchmark-findings branch starts at 46a1134, the
validated findings branch. Its own environment and structural index are isolated
from the running local-stress-20261010-r1 campaign. Main and campaign evidence
remain unchanged. Six new commits correspond to the six approved corrections. A separately
authorized seventh commit fixes silent conditional progress omissions.
| Correction | Scope | State |
|---|---|---|
| 1 | Shared CLI/MCP/daemon/family coverage classification | Implemented; focused tests and quality preflight pass |
| 2 | Clear terminal outcomes and timestamped progress | Implemented; 69 focused tests and quality preflight pass |
| 3 | Explicit renderer and outer architecture timeout controls | Implemented; 75 focused tests and quality preflight pass |
| 4 | Actual FAISS SIMD provenance and retained, quieter diagnostics | Implemented; 91 focused tests and quality preflight pass |
| 5 | Embedding phase attribution and interruption recovery | Implemented; 96 focused tests and final quality preflight pass |
| 6 | Integration, package/catalog/documentation alignment and full validation | Implemented; full gate passes: 1673 tests, 3 skips, 86% coverage |
| 7 | Explicit conditional skips without ordinal or resume fabrication | Implemented; 80 focused tests, quality preflight and full gate pass |
Coverage classification¶
The original MCP classifier checked failed selected files (partial) but omitted
broader incomplete coverage (complete=false). Thus usable indexes with zero
failed files could enter the complete-index cohort despite known coverage gaps.
CLI and MCP now share the predicate; ready daemon generation records also account
for coverage_complete=false. MCP/family aggregation checks all tool responses,
so a later complete response cannot erase an earlier incomplete member.
Operational errors and unusable indexes retain precedence. Empty query responses remain distinct from partial coverage and do not prove repository-wide absence. Timing eligibility is explicit for complete, partial and unusable outcomes. Resume regressions include failure-free coverage gaps as well as failed files.
The first focused validation passed 59 tests. Reclassification of 22 retained
MCP observations changed 19 original passed outcomes to partial; three were
already partial. This is a diagnostic replay, not a replacement campaign result.
Original receipts and timings were never overwritten. The replay with source
receipt digests is in the dedicated worktree's ignored
.artifacts/validation/stress-findings-20261013/coverage-reclassification.json.
The final quality preflight passes lint, formatting, typing, Semgrep, refreshed
indexing and docstring checks; initial typing failure evidence is retained.
Integration onto main and operator-run measurement campaigns follow completion of the current stress run. Existing malformed fixtures remain evidence; the runner does not add exclusions or retry deterministic failures automatically.
Terminal reporting¶
Timestamp behavior from eee8bef is incorporated into this correction without
altering the original branch. Stage ordinals and pending counts still exclude
retained resume checkpoints. Machine statuses remain unchanged; terminal text
explains partial coverage, known failed-file/coverage-issue counts, no matches,
real command failures and optional SVG warnings. Unknown counts are not invented.
An embedding-only response's empty issue list does not erase structural gaps.
Plain-text workspace reads receive coverage annotation before printing, so the
terminal and persisted receipts agree. The shared implementation covers both
local suites; successful protocol requests do not imply semantic completeness.
Architecture budgets¶
The runner previously imposed a 300-second outer deadline while silently using Codira's 60-second SVG default. Both suites now pass an explicit inner Graphviz limit, defaulting to the findings branch's 180 seconds. The outer default remains 300 seconds and includes extraction, report writing and rendering. Independent zero values disable only their respective limit; shorter outer limits remain valid operator controls and can interrupt rendering earlier. Both controls are printed in the dry plan and stored in immutable identity, so changing either requires a new run identity. Explicit per-command renderer arguments are retained.
Each successful repeated command preserves its renderer manifest beside raw stdout/stderr. Mandatory reports and optional SVG availability/timings are separate receipt fields; an optional timeout is shown prominently without being promoted to a command failure. Outer timeouts still terminate only the owned process group.
Native acceleration and MCP diagnostics¶
FAISS diagnostics record its version, binding module, compile options and, when available, dynamic dispatch and active SIMD level. A generic binding name does not imply generic execution: the inspected FAISS 1.15.0 wheel reports dynamic dispatch and active AVX2 despite lacking a separate AVX2 Python module. Reading these diagnostics never changes dispatch. Older wheels report unknown optional fields rather than guessing. The plugin distribution advances to 2.0.1 and the FAISS optional official bundle pin follows it in bundle 2.0.12.
Both benchmark suites retain MCP server stderr in each attempt's mcp.stderr,
referenced by its result receipt. Routine SDK/FAISS INFO messages no longer
interleave progress output. Actual tool errors still classify the attempt as
failed and name the affected tools; responses and attempt history remain intact.
The transport tests simulate the SDK boundary; native diagnostic tests inspect
the installed library. Neither is a corpus campaign or retrieval-quality claim.
Embedding population and interruption recovery¶
Both backends previously fetched the entire pending queue, performed inference for all selected texts, and committed structural pending removals only at the end. They now select bounded segments using the existing embedding work-size control and commit completed segments only for owned connections. Failed active segments roll back; borrowed transactions remain caller-owned. Explicit embeddings-only recovery also rebuilds similarity artifacts for an empty queue, closing the gap between vector persistence and interrupted publication. No schema or vector format changes are needed. SQLite advances to 2.0.3, DuckDB to 2.0.5 and their official bundle pins follow in 2.0.13.
Opt-in shared profiling separates queue selection, cache lookup, inference, serialization, persistence and commits, retaining interrupted phase evidence, throughput and correctly scoped lifetime RSS. Native database regressions use simulated inference; they verify both owned checkpoint recovery and borrowed rollback, plus explicit empty-queue publication and unavailable profiling. The population recovery guide documents commands, units, transaction limits and measurement scope. Original campaign timings do not isolate the slow phase; no whole-campaign improvement is claimed from the focused persistence probe.
Integration and final validation¶
The worktree inherits the validated findings at 46a1134, including canonical
method inventories/graph aliases, C++ header routing and measured DuckDB inventory
query repair. Timestamp changes are incorporated by correction 2. The final
package pins and generated installer catalog are aligned; README developer
possibilities link the benchmark and embedding recovery procedures. Rejected
DuckDB performance candidates remain reverted. Prior native measurements do not
qualify the original campaign or these new runtime/runner controls.
Corrections 1–5 have their focused tests and quality preflights. Correction 6 passed the combined full repository gate in a named tmux session before delivery: 1673 tests passed, 3 skipped, 86% coverage, including the seventh progress omission correction. That source patch is tested in the combined working tree and will be recorded after correction 6, preserving one commit per approved problem. Main, the original running checkout and campaign artifacts remain unchanged.
Conditional progress omissions¶
StageProgress.omit previously removed conditional stages silently. MCP probes
without a discovered identity could remove mcp/cli-evidence, making the pending
count jump with no corresponding terminal outcome. Newly omitted exact stages
and grouped children now each print one timestamped skip with their reason.
They decrement pending once and never advance measured-test ordinals. Repeated
omissions and retained resume checkpoints remain silent. No child measurement
receipt is fabricated for an omitted evidence probe.
Missing identities and embedding/concurrency prerequisites are exclusions; an actual MCP failure remains a failed measured probe and its dependent evidence is explicitly excluded. Both suites use this shared behavior, including final orchestration cleanup. Focused tests cover missing MCP identity, true tool errors, group counts, repeated omissions and retained partial checkpoint resume. The seventh commit follows the original six. Original campaign artifacts remain unchanged.