Bind entity-bound vars in pattern value position; widen Instant to int64 with cljs transit reps; fix multival 2-vector disambiguation

[tiensonqin]

Summary

Four engine/storage bugs hit by the logseq db-worker port.

Bug A + B — bound vars dropped in pattern value position (one root cause). When a variable bound to an entity id lands in a pattern's value position — (parent ?p ?c) inside a recursive rule body, or :in ?x where ?x was passed as Arg_scalar (Result_entity e) — the constraint was silently discarded. query_value_term only handled QValue; a var bound to an entity substitutes to QEntity, which fell through to None, so pattern_datoms's v-probe scanned every datom of the attribute and the relation fast path returned children of every parent. Now:

let query_value_term = function
  | QValue value -> Some value
  | QEntity entity_id -> Some (Ref entity_id)  (* a bound entity IS the ref value *)
  | _ -> None

(parent 1 ?c) over two disjoint trees now returns {2,3} instead of {2,3,6,7}; Arg_scalar (Result_entity 1) to :in ?x binds identically to Result_value (Ref 1).

Bug C — Instant/Uuid transit encoding incompatible with cljs. melange/datascript_melange_storage.ml (and its near-identical copies in sqlite/datascript_sqlite_codec.ml and examples/logseq_sqlite_storage.ml) wrote Transit.Tagged ("m", Transit.Int v) / Tagged ("u", String v):

• transit-js emitEncoded calls getVerboseHandler() for tagged reps under preferStrings → TypeError crash on every Instant write (bootstrap transact stores :db/txInstant, file/created-at);
• wire format ~#m/~#u tagged maps is not cljs's ~t/~u, so kvs blobs aren't byte-compatible;
• Instant of int truncates epoch ms on melange where int is 32-bit.

Fix: Ds.Instant widened to int64 (~30 use sites updated: parse_instant_millis now computes in Int64, murmur3_hash_long takes the int64 directly, Int64.to_string for display, `Intlit in the JS JSON binding), and all three codecs write Transit.Date value / Transit.Uuid value — the transit lib's cljs-compatible reps emitting ~t<iso> verbose / ~m<ms> normal and ~u. Legacy ~#m/~#u tagged blobs still decode via the kept read arms.

Bug D — a 2-vector in a cardinality-many ref attr was always treated as a lookup ref. ref_lookup_collection_value classified ANY 2-element List/Vector headed by an attr name as a lookup ref, so {:block/tags [:logseq.class/Page :logseq.class/Page]} raised Lookup ref attribute should be marked as :db/unique. Upstream maybe-wrap-multival (db.cljc) only keeps a 2-element collection as a lookup ref when (first vs) names a :db.unique/identity attr; otherwise it expands into individual values. Now:

let ref_lookup_collection_value context db = function
  | List [ attr; _ ] | Vector [ attr; _ ] ->
    (match attr_name_of_value attr with
     | Some attr -> context.is_unique_identity db attr  (* was: entity_ref_of_ref_attr_value value <> None *)
     | None -> false)
  | _ -> false

[:logseq.class/Page :logseq.class/Task] now expands into two ident-ref datoms (idempotent adds collapse); [:block/uuid u] still resolves as a lookup ref in both single- and multi-value attrs.

Schema validation regression fix (main's 1013dcf): the per-attr mid-tx refresh_schema now sees a partially-installed schema spec (db.type/tuple lands before db/tupleTypes), so validate_schema raised inside schema_from_transaction_datoms. schema_from_transaction_datoms gained ?(validate = true); the mid-tx refresh_schema calls it with ~validate:false (upstream never revalidates the whole schema mid-transaction), while the strict end-of-tx recompute still validates.

CI: workflow moved to OCaml 5.5.1 (required by lg/melange), datascript-ocaml-lg.opam gained pin-depends for lg (logseq/lg#main) so opam install --deps-only can resolve it, and the opam switch is cached keyed on OS/compiler/opam hashes.

Tests

test_query_engine_fixes:

• recursive_rule_bound_head_arg — logseq's (parent ?p ?c) rule over disjoint trees, literal and :in-bound head arg (both returned children of every parent before);
• in_scalar_result_entity — entity vs ref scalar inputs bind identically;
• many_ref_vector_of_idents_expands / many_ref_vector_of_same_ident_idempotent — vector of ident keywords in a many-ref attr expands into one datom per ident (dup collapses);
• many_ref_vector_with_unique_head_is_lookup_ref / entity_map_lookup_ref_vector_form — [:block/uuid u] still resolves as a lookup ref in One and Many attrs.

All verified fail-before/pass-after (Invalid_argument "Lookup ref attribute should be marked as :db/unique: [:logseq.class/Page :logseq.class/Task]" before).

test_sqlite_storage: sqlite_codec_instant_uuid_transit_reps asserts Instant/Uuid encode as Transit.Date/Transit.Uuid (verbose wire bytes contain ~t/~u), round-trip a >int32 millisecond value exactly, and still decode legacy ~#m/~#u blobs.

Full dune runtest green, including upstream cljs parity (identical datom hash) and the js_of_ocaml cross-runtime check.

Link to Devin session: https://app.devin.ai/sessions/c82925b6976a4a6da2688c410e65011a
Open in Devin Desktop: https://app.devin.ai/desktop/session/c82925b6976a4a6da2688c410e65011a?variant=devin
Requested by: @tiensonqin

[devin-ai-integration]

I'll fix CI failures and address comments from users with write access. I'll skip comments containing "(aside)".

• Disable automatic comment, CI, and merge conflict monitoring

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[devin-ai-integration]

Note on CI: the OCaml 5.2.1 check fails at opam install . --deps-only --with-test — datascript-ocaml-lg requires the lg package, which is unknown to the solver. That dep was added in 4c1ea15 (already on main; the same failure is on PR16). It's unrelated to this change — two things would be needed to fix it: a pin-depends for lg (git+https://github.com/logseq/lg.git), and lg requires ocaml >= 5.5 + melange >= 6.0, while the CI switch is on 5.2.1.

Locally on OCaml 5.5.1: dune build clean, full dune runtest green including upstream cljs parity.