Fix pull attribute ordering

[maxweber]

Summary

Two commits: the pull-attribute-ordering fix, plus the fixes for everything a high-effort multi-agent review of that commit confirmed.

cd8becb — Fix pull attribute ordering

dbval's indexes order attribute components by their serialized (pr-str) form, but pull's merge-join between the sorted pattern attributes and the :eavt datom stream compared attributes with keyword compare. The two orders disagree as soon as a pattern mixes plain and namespaced keywords (e.g. :type vs :instagram.story.image.preview/blob), desyncing the walk and silently dropping attributes from pull results. The fix centralizes the ordering in attr-sort-key/attr-compare and uses it in tuple-list, both pull comparison sites, and the parser's attribute sort (:db/id deliberately last). With regression test.

26dd323 — review findings, all confirmed by adversarial verification

1. FilteredDB pulls leaked other entities' data: attrs-frame's non-DB branch had its take-while outside the ->> — dead code under util/cond+ — leaving the scan unbounded and entity-unguarded. Pulling an attribute an entity lacks assoc'd the next entity's value into the result (pre-existing bug; the ordering commit edited the dead form without noticing). The take-while now lives inside the pipeline.
2. String-spelled keyword attributes silently dropped: \":name\" and \":db/id\" in a pattern — forms parse-attr-spec explicitly accepts — no longer matched keyword datoms. parse-attr-name now normalizes \":foo\" strings to keywords up front; bare strings still name genuine string attributes.
3. UTF-16 vs UTF-8 ordering divergence: attr-compare used Java string compare (UTF-16 code units) while the store orders serialized attributes by UTF-8 bytes (code points) — the orders disagree for supplementary-plane characters (emoji). Comparison now runs in code-point order via compare-attr-keys. Fixing this surfaced one more spot the review hadn't isolated: the parser's sort-by also used default string compare, inverting first-attr/last-attr for such patterns and making the pull return nil outright.
4. Hot-loop cleanup: the attribute sort key is computed once at parse time and cached on PullAttr instead of pr-str-ing the pattern attr's name on every frame-loop iteration.

Test plan

• New regression tests: namespaced-attr-after-ref-attr ordering (the original bug), FilteredDB bounded scan / no cross-entity leak, string-spelled attrs (\":name\", \":db/id\"), supplementary-plane attribute pairs.
• Full suite: 171 tests, 1075 assertions, 0 failures, 0 errors.

🤖 Generated with Claude Code

https://claude.ai/code/session_014dX8tTR4yFh5SyBGw3atpo