feat: goldens, honesty harness, SCIP ingest, multi-field vectors, native-only embed by AdityaVG13 · Pull Request #30 · AdityaVG13/ast-sgrep

AdityaVG13 · 2026-08-14T09:33:06Z

Summary

Campaign branch from the golden-assert testkit workstream, landed as regular commits (not squashed). Highlights:

Goldens / testkit (nz7i): assert_golden + scrubber, frozen CLI/MCP/handbook/extraction/chain dumps, compare-only golden CI + update SOP.
Honesty / conformance (1vhy, ghiw): default-on keep-gate vs committed bench history, FeatureUniverse matrix, oracle dispatch SSoT, Pass/Fail/Not-run reports, native pattern: vs ast-grep Pattern-1, tagged published bench ledgers.
IVF / fusion docs (hoy3, e2hc): measured floors, ANN threshold cliff (default unchanged), T1-R sidecar non-identity, fusion intern + stage budget scorecard.
Mock-free surfaces (lbx1.7-.12): watch daemon e2e, LSP stdio JSON-RPC e2e, Code Mode hashed embed-on, opt-in ast-grep --json spawn, ANN IVF scale tradeoff hard-fail on workflow_dispatch.
Bounds (1pi0, df74, pb2w): pattern walk file-size cap, Code Mode NAPI waiter abort, Pi symlink-out-of-root refuse, LSP indexPath fenced unless opt-in, wrapping SQLite byte offsets rejected.
Multi-field vectors (7d5x.1-.3): chunk body before metadata, per-field blobs + IVF fingerprint bind, query-intent field weighting. Eval recall on a frozen corpus is still open (7d5x.4).
SCIP ingest (kgvi.1-.3): fail-open load, matching defs/refs upgrade to ScipExact, asgrep index|reindex --scip PATH degrades never fails. Eval-pack precision-by-tier is still open (kgvi.4).
EmbedBackend cascade (f4ce.1-.3): SearchOptions adapters, CLI flags, LSP init options. Concurrent backends collapse Neural > Semantic > Auto.
HTTP embed removed: Cloud and Ollama clients, flags, env, and ureq are gone. Embeddings are in-process only (hashed default, optional ONNX neural). Stored cloud/ollama indexes fail closed until asgrep reindex. Cloudflare Code Mode adapter is unrelated and stays. lbx1.1 / lbx1.2 withdrawn.

Test plan

Targeted crate tests for the later landing commits (field_weight, scip/resolution, CLI search_options_ + machine_contracts with known skips, LSP embed_cascade + lsp_stdio_e2e), via rch exec.
HTTP-strip tests via rch exec: ast-sgrep-embed --lib, core embed_backend, CLI search_options_, LSP embed_cascade, core e2e_smoke parity, machine_contracts (known skips), MCP protocol (19 passed).
CI on this PR (full workspace / goldens / machine contracts).
Do not treat chain_expand_sample_dump_matches_golden float drift as in-scope (pre-existing).
Do not quote MRR/latency from this branch; eval beads 7d5x.4 / kgvi.4 are still open.
Neural default-on is not this PR; built-in hashed embeddings plus repository vocabulary remain the no-download default.

Adversarial review remediation

Confined golden updates and indexed evidence to canonical roots, invalidated stale schemas, corrected SCIP occurrence naming, and preserved multiline definition starts.
Removed a redundant tracked file from a gitignored raw-profile directory, stripped personal corpus paths, and marked measurements without retained raw evidence and corpus snapshots UNREPRODUCIBLE.
Review fixes: 0e4e284f, b555b1fc. CI was not awaited.

Co-authored-by: Cursor <cursoragent@cursor.com>

Seed intentional divergences and verdict conventions so XFAIL is never silent green, and stop claiming ast-grep CLI delegation. Co-authored-by: Cursor <cursoragent@cursor.com>

Install three pillar ledgers with Open pointer imports only, bridge fail-closed product cases, and mandate pre-flight mining in AGENTS.md. Co-authored-by: Cursor <cursoragent@cursor.com>

Replace the optional 50% tripwire with -3%/-5% class thresholds, cv quarantine, and host/SHA/profile attribution against .bench-history SSoT. Co-authored-by: Cursor <cursoragent@cursor.com>

Map each channel to an authoritative oracle and gate class so latency and unreproducible ledgers cannot be read as correctness. Co-authored-by: Cursor <cursoragent@cursor.com>

Encode present/partial/missing/excluded per host, keep MCP non-fusion as excluded, and leave conformal scoring to WP6. Co-authored-by: Cursor <cursoragent@cursor.com>

Hot-path hit fusion now interned u32 identity keys so multi-channel merge no longer clones file/symbol/caller/callee on every HashMap probe. Co-authored-by: Cursor <cursoragent@cursor.com>

Exclusive samply on a 403-file IVF-on cold index: extract 48.68%, blake3 9.77%, sqlite upsert 9.37%, ivf_build 7.56%. T3/UPSERT reopen_gate true pending Score>=2 and human review. Co-authored-by: Cursor <cursoragent@cursor.com>

Land living clause IDs for query grammar, machine envelope, and fail-closed negatives with table-driven parse tests. Score remains TBD; compact provenance drop stays a DISC, not a bug. Co-authored-by: Cursor <cursoragent@cursor.com>

Land the supported/unsupported matrix and an env-gated match-set harness. Default CI runs native hits plus fail-closed unsupported shapes; equality vs ast-grep stays Not-run without ASGREP_DIFF_AST_GREP. Co-authored-by: Cursor <cursoragent@cursor.com>

Per-section canonical/historical/UNREPRODUCIBLE/reproducible-in-tree tags. Delete dangling harness reproduce blocks. Archive the 110-file 285ms budget as breached. Quality snapshot stays UNREPRODUCIBLE; no new MRR invented. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Registry-driven proof-pack gate writes a report even on failure and stops overclaiming Pi release as local-release-gate.sh. Co-authored-by: Cursor <cursoragent@cursor.com>

Existing ASGREP_PERF_PROFILE events already split the stages; serial upsert is 65.6% of cold-index wall on this host. Co-authored-by: Cursor <cursoragent@cursor.com>

Live 1.965 s does not replace C1 4.304 s. No baselines.md floor rewrite. Co-authored-by: Cursor <cursoragent@cursor.com>

…(hoy3.4) IVF-on vs IVF-off at 2199 chunks is ~95 ms on this host. DEFAULT_ANN_THRESHOLD stays 2000. Co-authored-by: Cursor <cursoragent@cursor.com>

…3.5) Unit-dot simsimd is not bit-identical to f64 cosine. Same-metric kmeans reduce still is. No product change. Co-authored-by: Cursor <cursoragent@cursor.com>

Sub-1ms is the in-process CORE_PARTS gate, not CLI vs competitors. Ripgrep and ast-grep still win the small self-corpus CLI rows. Co-authored-by: Cursor <cursoragent@cursor.com>

Lower bound is 0. Present-count is not certified. No release_certificate.json. Co-authored-by: Cursor <cursoragent@cursor.com>

Keep the 2048+10000 test ignored on PRs. Recall@10 >= 0.99 stays the SLO; host µs remain UNREPRODUCIBLE. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Store name/docs/body/graph embeddings in SQLite (schema 10) and bump the IVF field-layout fingerprint so stale sidecars rebuild. Search still uses the concatenated primary vector until query weighting lands. Co-authored-by: Cursor <cursoragent@cursor.com>

Co-authored-by: Cursor <cursoragent@cursor.com>

Embeddings are in-process only: hashed default plus optional ONNX neural. Stored cloud/ollama indexes fail closed until reindex. Co-authored-by: Cursor <cursoragent@cursor.com>

Amp-Thread-ID: https://ampcode.com/threads/T-01a003f3-e62e-75ca-950f-0de88d7fe0fe Co-authored-by: Amp <amp@ampcode.com>

Amp-Thread-ID: https://ampcode.com/threads/T-01a006d4-a8ae-77c7-a203-288a069f14d5 Co-authored-by: Amp <amp@ampcode.com>

AdityaVG13 and others added 30 commits August 13, 2026 16:30

feat: add testkit assert_golden and scrubber (nz7i.1)

7d4b16f

Co-authored-by: Cursor <cursoragent@cursor.com>

docs: register DISC/COVERAGE harness shell (ghiw.1)

0da742d

Seed intentional divergences and verdict conventions so XFAIL is never silent green, and stop claiming ast-grep CLI delegation. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: add campaign negative ledgers under docs/progress (1vhy.3)

d48b480

Install three pillar ledgers with Open pointer imports only, bridge fail-closed product cases, and mandate pre-flight mining in AGENTS.md. Co-authored-by: Cursor <cursoragent@cursor.com>

feat: default-on keep-gate vs committed bench history (1vhy.1)

a790905

Replace the optional 50% tripwire with -3%/-5% class thresholds, cv quarantine, and host/SHA/profile attribution against .bench-history SSoT. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: add composite oracle dispatch SSoT (1vhy.4)

355d88b

Map each channel to an authoritative oracle and gate class so latency and unreproducible ledgers cannot be read as correctness. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: formal FeatureUniverse surface matrix (1vhy.5)

7e5a53f

Encode present/partial/missing/excluded per host, keep MCP non-fusion as excluded, and leave conformal scoring to WP6. Co-authored-by: Cursor <cursoragent@cursor.com>

feat: intern fusion dedup keys instead of cloning strings (e2hc.30)

8038346

Hot-path hit fusion now interned u32 identity keys so multi-channel merge no longer clones file/symbol/caller/callee on every HashMap probe. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: record post-T1R residual leaf shares (hoy3.1)

a953d37

Exclusive samply on a 403-file IVF-on cold index: extract 48.68%, blake3 9.77%, sqlite upsert 9.37%, ivf_build 7.56%. T3/UPSERT reopen_gate true pending Score>=2 and human review. Co-authored-by: Cursor <cursoragent@cursor.com>

feat: number QG/MJ/NL MUST clauses (ghiw.2)

4a03c08

Land living clause IDs for query grammar, machine envelope, and fail-closed negatives with table-driven parse tests. Score remains TBD; compact provenance drop stays a DISC, not a bug. Co-authored-by: Cursor <cursoragent@cursor.com>

feat: freeze CLI search dumps and teaching messages (nz7i.2)

85d0243

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: freeze handbook, MCP tools/list, and catalog dumps (nz7i.3)

c97594f

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: freeze extraction dumps and chain expand JSON (nz7i.4)

ba93fcf

Co-authored-by: Cursor <cursoragent@cursor.com>

ci: pin compare-only goldens and document update SOP (nz7i.5)

d5be2e5

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: add fixture provenance and IVF/migration corpora (ghiw.4)

c455c69

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: emit Pass/Fail/Not-run compliance reports (ghiw.5)

9be8d52

Registry-driven proof-pack gate writes a report even on failure and stops overclaiming Pi release as local-release-gate.sh. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: record S2 stage wall timers for prepare vs upsert vs IVF (hoy3.2)

ac6bcf1

Existing ASGREP_PERF_PROFILE events already split the stages; serial upsert is 65.6% of cold-index wall on this host. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: add S2 MEASURED FLOOR promotion MATCH protocol (hoy3.3)

0c5e83a

Live 1.965 s does not replace C1 4.304 s. No baselines.md floor rewrite. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: measure post-T1-R ANN threshold cliff without changing default …

a836638

…(hoy3.4) IVF-on vs IVF-off at 2199 chunks is ~95 ms on this host. DEFAULT_ANN_THRESHOLD stays 2000. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: record T1-R IVF sidecar non-identity vs pre-T1 cosine path (hoy…

41a4189

…3.5) Unit-dot simsimd is not bit-identical to f64 cosine. Same-metric kmeans reduce still is. No product change. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: publish honest fusion stage budget scorecard (e2hc.31)

725599c

Sub-1ms is the in-process CORE_PARTS gate, not CLI vs competitors. Ripgrep and ast-grep still win the small self-corpus CLI rows. Co-authored-by: Cursor <cursoragent@cursor.com>

docs: seed red conformal score and multi-ref cert checklist (1vhy.6)

1c8a3c5

Lower bound is 0. Present-count is not certified. No release_certificate.json. Co-authored-by: Cursor <cursoragent@cursor.com>

ci: hard-fail ANN IVF scale tradeoff on workflow_dispatch (lbx1.7)

b15ad78

Keep the 2048+10000 test ignored on PRs. Recall@10 >= 0.99 stays the SLO; host µs remain UNREPRODUCIBLE. Co-authored-by: Cursor <cursoragent@cursor.com>

test: add real CLI watch daemon e2e (lbx1.8)

4409101

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: parse opt-in ast-grep --json spawn (lbx1.9)

f39ea6b

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: MCP semantic_search hard-fails unconfigured cloud (lbx1.10)

ab85649

Co-authored-by: Cursor <cursoragent@cursor.com>

test: cover codemode session embed-on hashed path (lbx1.11)

c1227f8

Co-authored-by: Cursor <cursoragent@cursor.com>

test: add LSP stdio JSON-RPC process e2e (lbx1.12)

f500bec

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: fence LSP indexPath to workspace or cache unless opt-in (df74)

8488789

Co-authored-by: Cursor <cursoragent@cursor.com>

AdityaVG13 and others added 19 commits August 14, 2026 00:39

refactor: feature-gate testkit→lsp (e17)

e32f004

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor: root-reexport HitKind and related core DTOs (b1i5)

4177e55

Co-authored-by: Cursor <cursoragent@cursor.com>

docs: record post-E17 with-dev SCC Keep (sodf)

3f54272

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: cap native pattern walk at the index file-size bound (1pi0.1)

f533141

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: bound embed HTTP timeout and response body size (1pi0.2)

ad679b9

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: abort Code Mode NAPI waiters on soft timeout (1pi0.3)

e7dde22

Co-authored-by: Cursor <cursoragent@cursor.com>

test: refuse Pi edit through a symlink out of root (1pi0.4)

8ce27a9

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: reject wrapping SQLite byte offsets (pb2w)

a73a131

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: render semantic chunk body before metadata (7d5x.1)

970b4c6

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: fail-open SCIP JSON load without graph mutation (kgvi.1)

ee01164

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor: route SearchOptions embed flags through EmbedBackend (f4ce.1)

6919427

Co-authored-by: Cursor <cursoragent@cursor.com>

fix: bind IVF fingerprint to semantic field layout (7d5x.2.1)

d736781

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: weight embed fields by query intent (7d5x.3)

f945c0b

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: upgrade matching graph edges to ScipExact (kgvi.2)

cbf746f

Co-authored-by: Cursor <cursoragent@cursor.com>

feat: add asgrep index --scip with fail-open degrade (kgvi.3)

682e146

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor: cascade CLI embed flags through EmbedBackend (f4ce.2)

86b9adc

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor: cascade LSP embed settings through EmbedBackend (f4ce.3)

ab91c79

Co-authored-by: Cursor <cursoragent@cursor.com>

refactor: remove cloud and Ollama HTTP embed backends

7a004f1

Embeddings are in-process only: hashed default plus optional ONNX neural. Stored cloud/ollama indexes fail closed until reindex. Co-authored-by: Cursor <cursoragent@cursor.com>

AdityaVG13 changed the title ~~feat: goldens, honesty harness, SCIP ingest, multi-field vectors, embed cascade~~ feat: goldens, honesty harness, SCIP ingest, multi-field vectors, native-only embed Aug 14, 2026

AdityaVG13 mentioned this pull request Aug 15, 2026

feat: in-process critic, causal planner, and channel conjunction #31

Merged

AdityaVG13 and others added 4 commits August 15, 2026 05:27

feat: report graph precision by resolution tier

87629cf

Amp-Thread-ID: https://ampcode.com/threads/T-01a003f3-e62e-75ca-950f-0de88d7fe0fe Co-authored-by: Amp <amp@ampcode.com>

fix: harden indexed evidence contracts

0e4e284

Amp-Thread-ID: https://ampcode.com/threads/T-01a003f3-e62e-75ca-950f-0de88d7fe0fe Co-authored-by: Amp <amp@ampcode.com>

docs: label unretained benchmarks honestly

b555b1f

Amp-Thread-ID: https://ampcode.com/threads/T-01a003f3-e62e-75ca-950f-0de88d7fe0fe Co-authored-by: Amp <amp@ampcode.com>

Merge main into feat/golden-assert-testkit

aa68770

Amp-Thread-ID: https://ampcode.com/threads/T-01a006d4-a8ae-77c7-a203-288a069f14d5 Co-authored-by: Amp <amp@ampcode.com>

AdityaVG13 merged commit 38960f0 into main Aug 15, 2026
8 of 11 checks passed

AdityaVG13 deleted the feat/golden-assert-testkit branch August 15, 2026 23:01

Navigation Menu

feat: goldens, honesty harness, SCIP ingest, multi-field vectors, native-only embed - #30

feat: goldens, honesty harness, SCIP ingest, multi-field vectors, native-only embed#30
AdityaVG13 merged 53 commits into
mainfrom
feat/golden-assert-testkit

AdityaVG13 commented Aug 14, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

AdityaVG13 commented Aug 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Adversarial review remediation

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

AdityaVG13 commented Aug 14, 2026 •

edited

Loading