Release train 0.36.0: Ruby package owns its source, Rails DSL, Rust/Ruby recall, faster package start #352
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!352
Loading…
Reference in a new issue
No description provided.
Delete branch "train-0.36"
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 the v0.36 release train: one full local run, one PR, one CI round, one tag set. It closes #337, #338, #341, #345, #347, #348, #349 and #350.
What is in it
pkg/de.h-dv.ruby/v0.10.0).let/subjectnames shadow model methods of the same name. A DSL method is not emitted when its class already defines a method of that name.attr_readerresolves as a read.crate::counts as origin. This wins back correct binds that #336 dropped. Schema v81.daemon_restarting.promotion_resolving.package_recordskeeps a host-release row on the host release.ruby_package_cost_attributionattributes the Ruby cost record's own leg per statement.manifest_layeringattributes#[path]files to the crate that compiles them.index_stateblock is trimmed back under its ceiling.Records (all operator-approved, each with attribution in its reason)
baseline.json,stage-baseline.jsonandtier3-baseline.json: #337/#348 binds, source-sampled before the re-record.cost-baseline.json: the rollup lever's full scan (+6–11% fullscan) and the kind no-op fix (vm_step −0.4…−3.9%).ruby-package-cost.json: schema 81 / ruby 0.10.0, attributed per statement.Verification
Every lane was verified independently: its claimed tests were re-run, and a mutation of its fix turned them RED.
The ONE full local run on
805a38c0was all green:-D warnings --document-private-items;COSI_GUEST_REBUILD=1(none unmeasured), the release gates, the ruby corpus leg, and builtin overhead parity for all seven languages.Not run locally:
agent_task_bench, because the local rule forbidsCOSI_CORPUS_DIRon MCP tests. CI runs it.Release order after merge
pkg/de.h-dv.ruby/v0.10.0. This is the first package-release.yml run; it packs with the newest core release's binary.🤖 Generated with Claude Code
https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmu
`active_grant_gate("cd", "b", ..)` rendered `cd.generation_id = b.id`, comparing a generation id to a language_bridges ROW id. It now passes the already-joined active generation `g` (== b.generation_id). Test: influence_classification::the_bridge_destination_gate_keys_on_the_ generation_not_the_bridge_row, both directions, mutation RUN (RED). The I340 differential oracle now spells b.generation_id, and its bridge row id no longer coincides with the generation id, so it grades the missed direction too (mutation RUN: 36 refs differ). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuRoot cause: every graceful exit records its cause, closes its listener, then holds the daemon lock through its drain (accept drain, watcher stop, final WAL checkpoint) and removes its registration last. A daemon spawned in that window ran `takeover::acquire_with_takeover`, which read the leaving holder as healthy (probe answered before the listener closed), as STARTING (young, unreachable) or as payload-less, and DEFERRED: the spawn exited, the old daemon finished, and nothing served. The MCP attach then waited out its 60 s poll and fell back to single-process mode ("did not answer within 60s. OBSERVED: NEVER-SPAWNED"), reproduced by the new e2e under the mutation. A holder past the 60 s startup grace was instead EVICTED: SIGKILL mid-checkpoint. Fix (daemon side, so every spawner benefits): a holder whose exit record names its pid, written since it registered, is LEAVING; so is a held lock with no registration. The successor waits up to 30 s for the lock instead of deferring or evicting, then falls back to the old health decision. Client side: when a heal fails after the previous daemon announced its exit, the error is a typed `DaemonRestarting` naming the old pid and its cause, kept for every later RPC of the call, and the MCP server renders it as the retryable `daemon_restarting` instead of `internal_error`. Debug-only daemon seams (env, like CODE_INDEX_DAEMON_TEST_PANIC_WHEN): CODE_INDEX_DAEMON_TEST_IDLE_EXIT_WHEN (idle exit on a marker) and CODE_INDEX_DAEMON_TEST_SHUTDOWN_HOLD_MS (hold the lock after the listener closes). No #[cfg(test)] static was added. Tests (mutation RUN for each, all RED): - reconnect_e2e::a_daemon_started_during_an_idle_exit_waits_for_the_lock_and_serves (3 tries) - daemon_death_mcp_e2e::a_session_started_during_an_idle_exit_is_answered_by_a_new_daemon (2 tries) - daemon_death_mcp_e2e::a_heal_that_cannot_finish_after_an_idle_exit_answers_daemon_restarting - takeover unit tests: announced exit waited for, payload-less lock waited for, a pre-registration exit record ignored - server::call_timeout_render_tests::a_daemon_restart_in_progress_is_retryable_and_names_the_old_exit Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmupromotion_resolving(#341) ca090f2210A path with no `files` row is DEFERRED by an activation build: it carries no rows, and the first ordinary walk after promotion full-upserts it. The build nonetheless read, extracted and validated it first, and dropped every byte of the answer. On a FRESH database every path is such a path, so a package's first index extracted the whole tree twice — `index::reconcile_activation_before_index` builds and promotes an isolated generation, then the ordinary pass extracts again. `dispatch_round_observed` now settles an owed path that has no `files` row without reading it: `defer_one` stats it against the recorded identity (Vanished / Moved -> redispatch, exactly as before) and routes it with `classify_walked`, which reads no bytes (Routing -> unclaimed, as before); anything claimed settles `done` with no file_id and a NULL hash ("recorded from a stat alone", m0053). A `files` row that appears between dispatch and settlement leaves the path owed, so the next round extracts it like any indexed file. WHY THE #78 GATE IS UNCHANGED. Every predicate `build::gate` applies reads the queue's identity columns, this generation's file_contributions/symbols/refs/imports, or `files` joined to them. A path with no `files` row reaches none of them whatever its extraction answered: `dispatch_round` writes a refusal contribution only for a path WITH a `files` row, so before this change a trapping or lying extractor on a fresh database was not refused by the gate either — the build promoted, and the ORDINARY pass recorded one refusal per file. Both halves still hold and are graded: the routing and churn halves of the gate see the same rows, and `first_index_single_extraction.rs`'s broken-package arms pass on the OLD code at every gate-state and refusal assertion (promoted, no failure, one refusal contribution per file on the package's component, zero symbols, parse_errors == files). One text moved and it is named: the per-file refusal used to read `host.restart_backoff (...; last=host.worker_trapped)` because the package had already spent its failures on the discarded extraction; it now reads the extractor's own `host.worker_trapped` (`fact.span_out_of_range` for the lying guest). `PackageHost::extractions_answered` exposes the count the lane pool already kept (LaneGate::served), so a test can say how many times one pass asked a package for a file. Cold ratio, median of 5 paired rounds, all seven first-party packages installed, load < 4 at start, before (2aa240b7) -> after: python-flask 1.361 -> 1.342, ruby-sinatra 1.500 -> 1.357, php-guzzle 1.365 -> 1.200, rust-ripgrep 1.321 -> 1.191, cs-dapper 1.489 -> 1.344, js-express 1.427 -> 1.317, ts-zod 1.166 -> 1.131. Parity exact on all. Tests, each mutation RUN (cp snapshot, md5, touch): * first_index_single_extraction::a_first_index_extracts_each_claimed_file_exactly_once - build.rs as before this change: RED (6 requests for 3 files); the two broken arms RED only at their final request count; pass_policy::a_builtin_transition_preserves_the_retained_package_record RED (parse_count 3, not 2). * ..._a_trapping_package_on_a_fresh_database_is_refused_per_file_as_before and ..._an_extractor_emitting_invalid_facts_... - Deferred settled `unclaimed`: RED, ActivationFailed/gate_failed. - extract_for answers a refusal with empty facts: RED. * pass_policy's parse_count assertion moves 3 -> 2: it counted the double read this removes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuCI run 1066 (nextest profile `ci`, beside the suite) failed svelte_package_cost at its FLOOR: "the package leg was CHEAPER than the TypeScript alone (61%): it skipped work". Its own numbers say otherwise: builtin_ms=1041 package_ms=637, but package vm_step 4,209,040 against 3,157,220 and stmts 10,845 against 8,794 -- the package leg did 33% MORE SQLite work, and a neighbour had slowed the builtin leg. A wall ratio answers "did it skip work" only on an idle box. * The floor is now `p.vm_step >= b.vm_step && p.stmts >= b.stmts`: the package leg writes the same TypeScript rows plus a template's. The counters are the leg's own and identical locally and on CI (4,209,040 / 10,845). * The wall-clock CEILING (ratio <= 1500, the amplification bound) stays. It is an upper bound on measured wall time, so the test runs alone in `ci-timing`. It was never listed because the duration comes back from a helper (`work::measure` -> `WorkUnits::wall_ms`), which wall_clock_registry's body scan could not see. The registry now counts a `.wall_ms` field as it counts `.elapsed()`, which finds exactly the two tests that bound a ratio of them, and `.config/nextest.toml` lists both in `ci-timing` and in `ci`'s exclusion. * vue_package_cost had the same floor and the same ceiling (it says it is "the svelte suite's"); fixed the same way. Mutations (RUN; cp snapshot + md5 + touch): * svelte: `let p = WorkUnits { vm_step: b.vm_step - 1, ..p }` -> RED at the work floor. vue: the same with `stmts` -> RED at the work floor. * `embedded::delegate` returns no regions -> RED at the parity assertion in both (documented, re-run). * registry: `MEASURED` without `.wall_ms` -> RED (both terms stale, and the new anti-vacuity `take` case); svelte's term deleted from both profiles -> RED naming it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0126PDDLB4wNHxKXvWM1VNmuShipped. Head
6ec0a612is on master's first-parent line, because master was fast-forwarded to it rather than merged through this PR, and it is tagged v0.36.0. Closing as delivered.Pull request closed