No reviewers
Labels
No labels
code-review
correctness
dos
performance
security
severity/high
severity/low
severity/medium
tech-debt
Kind/Breaking
Kind/Bug
Kind/Documentation
Kind/Enhancement
Kind/Feature
Kind/Security
Kind/Testing
Priority
Critical
Priority
High
Priority
Low
Priority
Medium
Reviewed
Confirmed
Reviewed
Duplicate
Reviewed
Invalid
Reviewed
Won't Fix
Status
Abandoned
Status
Blocked
Status
Need More Info
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
h-dv/code-index!351
Loading…
Reference in a new issue
No description provided.
Delete branch "train-0.35"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
This is release train v0.35.0. The operator asked to batch issues into one CI round and one release instead of one round per issue. All lanes were integrated on
train-0.35, rebased onto master05dcbee2(#344 CI fast path), version-bumped to 0.35.0, and run through ONE full local gate run.Contents
<script>over.langjoins the resolver's group keys. Without it, one file holding two languages that use one name failed the WHOLE index pass withUNIQUE constraint failed: resolve_groups….<script>/<script setup>via embedded regions; template component tags, handlers, interpolations. On Zammad, 525.vuefiles gain 5,472 symbols and 7,331 newly resolved refs.extendsprofile; re-recording it is not a move.plugin statusnames a live daemon's build instead of "interrupted — retry".code-index://registry/pluginsis served whole or refused, never truncated (the catalog had outgrown the 4k cap). APss-reading test runs alone.disclosure_derivation_registryhandles braceless#[cfg(test)]items.Records re-recorded (each operator-approved, with categorised reasons)
::Loggerwrong bind is named).cost-baseline.jsonwhole-record re-measured after two reduction rounds.embedded_dispatch_versioncondition.baseline.jsonrows for repos that did not move are byte-identical.Local full run (train @
166f2a3a, before the fix-ups)-D warnings --document-private-items) andguest_gates.sh: clean.ci: 4816/4819. The 3 failures were fixed by the fix-ups and re-run green (registry_mcp_e2e7/7 in both legs,idle_worker_footprint).ci-timing: 31/31.fast_gates.sh62/62 targets, clippy, fmt.Known limits (in the release notes)
Note
The title of
c891fc75says "CORPUS RECORDS NOT RE-RECORDED".f8505460re-records them, and the history is kept as is.🤖 Generated with Claude Code
https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmu
A new optional fact record, 0x8009: `span | language u16 | scope u8`. The language is an INDEX into the host-owned EMBEDDED_LANGUAGES table (typescript, javascript), never a string; scope says whether the region is an instance script (nothing importable) or a module script. validate() refuses, each from one comparison: * end past the file / inverted fact.region_out_of_range * a region inside or around its predecessor (depth budget is one) fact.region_depth_exceeded * an index outside the table fact.region_language_unknown (new, appended) * a guest fact starting inside a region the same frame delegated fact.language_not_owned * empty, descending, partial overlap, scope byte > 1 abi.non_canonical * more than MAX_EMBEDDED_REGIONS (64) fact.count_exceeded Each extent is charged against max_decoded_bytes; disjointness bounds the parsed bytes by the file, so the "64 one-byte regions at EOF parse 64 copies of the file" shape is not reachable. Skippable, and skipping is thinner rather than wrong, so ABI_MINOR does not move. Three producerless reason codes gain producers; reason_producers' PRODUCERLESS and the reason-code register drop their rows in the same commit. The guest kit gains the language rows, the region tag/scope bytes and the ceiling as CONSTANTS ONLY: an SDK method moved all nine build.py extractors' bytes (measured), so the svelte guest builds the record through the existing Encoder::raw. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuA security review flagged a0e5768b for capability-grant bleed. Measured: * ensure_delegated_component COPIED its siblings' grants across every generation at mint time, and write_capability_grants then re-granted every component of the package. Each masked the other: removing the re-grant stayed GREEN because the copy covered it, and widening the copy into the current generation stayed GREEN because the re-grant replaced it — so neither was graded. The copy is removed: a delegated component is minted INERT and granted only by the two package-level writers (write_capability_grants on the ordinary pass, which runs after extraction and before the resolve; project_generation_grants per generation, joined by package, at build and promotion). Both REPLACE. tests/delegated_grant_boundary.rs, through the production doors: 1. its capabilities equal the package's through both -> one -> none -> both re-approvals; 2. a component minted after a narrowing (wide generation retained as superseded) gets the narrow grant; 3. a pass with no package drops template and script rows together; a declined package leaves no live delegated row and no active grant (counted on the LIVE projection — the superseded generation keeps its exclusive rows for rollback, which is not serving them); 4. no package reaches another's delegated component; a claim named `typescript` with another row language is refused; both reserved keys are refused. Each with a run mutation; results in the doc comments. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmu* lang_literal_gate — the host spelled two builtin language ids (the ABI's region table and `embedded::DELEGABLE`). The table now holds ROUTES, by file extension: `EMBEDDED_ROUTES = ["ts", "js"]` (`EmbeddedRouteId`, `Region::route`). The host dispatches a region as `embedded.<ext>` to the compiled-in set exactly as the walker would, and the row language is whatever the chosen extractor stamps — no language id in host code, no allowlist entry. Wire bytes unchanged (index 0/1), so the Svelte extractor and its digests do not move; its provenance re-records the one edited comment in region.rs. `the_route_table_is_pinned_beside_the_dispatch_version` pins the routes to the dispatch version and checks, in a test, which language each lands in. * ignored_test_reachability — the #[ignore]d per-statement attribution in svelte_package_cost is waived the way `cost_attribution` is: a diagnostic with no bound; the gated half of the same target runs in every workspace run. * bounding_site_registry — MAX_EMBEDDED_REGIONS registered with FieldIn("daemon", "parse_error_files + error"): it refuses, never truncates. New embedded_region_ceiling.rs proves the claim through a real worker and index pass: 64 regions accepted and all delegated, 65 refuse the whole file with `fact.count_exceeded` in parse_error and nothing stored. Mutations RUN, all RED: ceiling `>=`→`>`; a route appended; a builtin literal reintroduced in embedded.rs; the waiver's target renamed; the registry row renamed; a fabricated disclosure field. The embedded.rs suite's earlier mutations re-run after the change, all RED. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuEmbedded-region dispatch put a package's template rows and the host's delegated <script> rows under one file_id. A template ref and a script ref of one unresolved name (`onclick={inc}` beside `inc();`) then made two groups agreeing on (file_id, name, pool_class, qualified, recv_unproven), and the INSERT into temp.resolve_groups failed its PRIMARY KEY: the index pass exited 1, the daemon never reported ready, and nothing in the project resolved. `lang` (the ref's own row language) is now in the key of every group-keyed resolver table -- resolve_groups, resolve_stats, resolve_decisions, reachable_stats (both builders), t3_groups, t3_reach, qual_groups, qual_decisions, tqt_groups, tqt_decisions -- and every statement that correlates on a group key carries it: the tier 1a/1b/2/3 writers, the tier-3 estimate, ambiguous_remain, follow_own_imports, the shadow refusal, the per-tier attribution, the apply UPDATEs of tiers 1/2/3, 1Q and T, and the 1Q receiver veto. tier1q_rel_samefile, the one 1Q pass whose candidate pool was not lang-scoped (same file used to imply same language), now is. Each language's group is decided within its own language's pool, so the template's `inc` (de.h-dv.svelte/svelte) stays unresolved: the function is a JavaScript symbol, and a same-file cross-language tier is #331's work. Both new suites assert that ceiling so it goes red when it moves. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuTaken from train-0.35's records and re-blessed with the sanctioned switches only; no number was merged by hand. Each `_blessed` carries its reason and categorised diff, and each file was checked key by key against train-0.35's bytes. baseline.json (COSI_CORPUS_BLESS) ruby-sinatra row only: symbols 1260 -> 2800, refs +859, resolved 3070 -> 3559 (+489), edges 1173 -> 2909, resolved_type +210, resolved_member_access +279. The reason names the one known wrong bind among the +489 (lib/sinatra/base.rb:1844 `::Logger` -> Sinatra::Middleware::Logger, the host's pre-existing rule-11 `::X` bypass). stage-baseline.json (COSI_STAGE_BLESS) ruby-sinatra rules only. host-package-lifecycle.json (COSI_PACKAGE_BLESS) ruby/* states only; #275's re-recorded fields and every xaml/svelte state are untouched. What moves is the one BUILTIN-owned un-shadowed sinatra file (+8 symbols, +2 refs); the frozen 0.8.1 package does not. cost-baseline.json (COSI_COST_BLESS) WHOLE RECORD, operator-approved 2026-10-09: ruby-sinatra 15.75M -> 17.28M vm_step (+9.7%), every other repo's ceiling tightens (ripgrep -14.7%, guzzle -13.4%, the rest -3.3..-6.6%) after #335's two HOST reductions; schema stamp 78 -> 79. Load average 3.4 -> 3.2 across the run. ruby-package-cost.json (COSI_RUBY_PKG_COST_BLESS) package 0.9.0: package leg vm_step 29.34M -> 31.48M, builtin 16.31M -> 17.17M. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuA host that adds a trait made every stored declared profile unreadable ("profile.(fields) = 17 of 18"): load_book fell back to the builtin, silently dropping the package's overrides for migration and promotion passes, and the next ordinary pass re-wrote the row in full and called the new spelling a MOVE -- one spurious Full re-resolve per upgrade per package-owned project. - ProfileTraits::decode_over: a field the row does not carry takes the extends profile's value. A build that did not know a field also refused any manifest overriding it, so the extends default IS the package's answer. Unknown, repeated and unreadable fields are still refused. - profiles::load_book decodes each row over traits_of(wire id), which is the profile its declaration extends. - profiles::write_declared compares DECODED profiles, not stored strings. Tests (mutations RUN, each RED): lang_profile::an_older_row_decodes_over_its_extends_profile - decode_over from NONE; decode_over strict profiles::an_older_hosts_row_is_read_and_rewriting_it_is_no_move - decode_row strict; fingerprint over stored strings Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmudurable_state read the generation table alone, and a building row is the same row whether its process is alive or dead. So a daemon mid-build (generation 5, 6,518 of 10,208 files) was reported as 'interrupted mid-build -- code-index plugin retry', inviting a second activation against a live one. The owner is already measured -- activation::owner, the lockfile (incl. daemon-<pid>.toml) plus port-connect probe every mutating command defers on. cmd_status now probes it ONCE and hands it to durable_state, whose owed line (activation::owed_line) says: - live daemon: 'building generation N (owner: the running daemon, pid P, port X, D of T queued file(s) settled) -- the owner drives this generation itself', no retry advice; - no live owner: the existing interrupted wording and retry repair; - undetermined owner: says so, and conditions the repair on it. Tests (mutations RUN, each RED): plugin_activation_cli::status_tells_a_live_daemons_build_from_an_interrupted_one (real building generation + build queue; dead owner, then a real listener behind a payload) - daemon arm rendered as Cli; cmd_status passing Owner::Cli; Cli arm rendered as the daemon arm activation::tests::an_undetermined_owner_and_an_idle_index_say_what_they_measured - Unknown rendered as Cli; unreadable progress dropped Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuStep 3 ("COMPETED") of `apply_general` spliced a correlated competitor subquery into its UPDATE twice (EXISTS witness + MIN in SET), probing every same-named symbol for every ref in the project. It was 89% of the Vue package's extra SQLite work on Zammad and 86% of Svelte's. Now the relation is built ONCE per pass into temp.influence_competed (ref_id PRIMARY KEY, MIN component) and the UPDATE is keyed on it: * UNSCOPED pass (cold, or after a grant change): over temp.influence_dyn_names, the admitted dynamic symbols folded to one row per (name, capability, component, file-or-0) with n and MIN(id). "some candidate is not the target" == n > 1 OR sid <> target. Only the same-file channel keeps rows per file; a unit test holds every `reaches` to reading no symbol column but s.file_id. * SCOPED pass: the symbol-level join driven from the in-scope refs, so a one-file edit pays for its scope only. The anchor arm's three qualifier channels are decided once per DISTINCT qualifier into temp.influence_anchored_quals and joined back by equality, the shape path_evidence_match_sql uses. Results identical. Measured, vm_step (load avg 6-16 throughout): * Zammad + de.h-dv.vue, cold: step 3 516.1M -> 35.3M; whole index 1,728.3M -> 1,247.7M (-27.8%). All 684,900 refs' (influence, component) byte-identical to the old code (157,574 influenced). A symbol-level dynamic-driven join was tried first: 398.1M. * svelte_package_cost: package leg 7,372,206 -> 4,506,497; step 3 3,392,810 -> ~0.50M. vue_package_cost: 7,494,503 -> 4,647,516. * cost_attribution, nine pinned repos: influence statements net 0 on every repo (the short-circuit runs there); totals move +-0.03% in FTS segment merges only. influence_step3_differential: the pre-#340 pass written out as an independent literal-SQL oracle (own channel table, capability names, fold and bridge gates), validated against the ORIGINAL code first, then compared row-for-row by refs.id under unscoped, scoped and batch scopes from a seeded mix of classes, with population floors asserted. Twelve mutations run and recorded at the test; MIN->MAX (both shapes), the count test, the same-file fold, the anchor prefix flip and the separator check all SURVIVE influence_classification and go red here. ref_kind_stance_registry: competitors and competitors_by_name now read `FROM refs` and are declared, with apply_general's stance. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmu77976513under schema 80 (#336, #340; operator-approved 2026-10-09) 74392dbb86The resource was capped by READ_CODE_MAX_TOKENS (4,000) and the catalog crossed it — the de.h-dv.vue row plus the bumped svelte/ruby rows took registry.v1.json from 13,167 to 14,399 bytes — so clients got `{"truncated":true,"preview":…}`: half a JSON catalog that parses as nothing. registry_mcp_e2e's two catalog tests went red on it. It now gets the skill-file arm's failure mode under its OWN stated ceiling, REGISTRY_RESOURCE_MAX_TOKENS (12,000, envelope included): served whole, or refused with an error naming its size and the cap. `the_registry_resource_fits_its_ceiling_with_the_worst_case_matches` keeps the refusal unreachable for the embedded catalog: it renders the catalog with every claimed extension symbol-blind and every package installed (27 match rows for 11 packages) and grades it through `headroom::Ceiling` — 5,209 of 12,000 tokens, with 4,000 reserved for the envelope (REGISTRY_BODY_BUDGET_TOKENS = 8,000). MUTATION (RUN): body budget = READ_CODE_MAX_TOKENS, the old cap -> RED. Both constants are registered in bounding_site_registry; the ceiling is graded through the headroom mechanism (payload_headroom_registry). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmu;`production_server()` strips every `#[cfg(test)]` item by skipping to the first `{` and its matching `}`. A braceless item — the test-only `#[cfg(test)] const REGISTRY_BODY_BUDGET_TOKENS` added in 360d9f8b — has no `{`, so the skip ran on through the NEXT item and deleted the production `registry_plugins_resource`; the gate then reported that surface as "declared and call NO renderer any more". The skip now stops at a braceless item's `;`. MUTATION (RUN): without the new arm, on this tree -> RED, `the_surface_set_is_derived_not_asserted`: ["registry_plugins_resource"]. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuFast-forwarded into master at
bd683200and released as v0.35.0 in release run 1062, with 60 assets: https://git.h-dv.de/h-dv/code-index/releases/tag/v0.35.0CI was green on the PR (runs 1056–1058) and on the master push (runs 1059–1061). Forgejo does not mark a fast-forward as merged, so I'm closing this by hand.
.nameon an untyped receiver binds "resolved" to the only in-projectdef name(phantom edges into context_pack / change_impact / safe_delete) #336plugin statusreports a live daemon build as "interrupted mid-build — run plugin retry" #346Pull request closed