test: ObservableQuery equivalence fuzzer (+ fix stale result on coalesced delete+recreate)

[danReynolds]

Summary

Adds an equivalence fuzzer for ObservableQuery and fixes a stale-result bug it found.

ObservableQuery maintains its result incrementally — on each broadcast it inspects only the changed documents and patches a cached result rather than recomputing. The property under test: the value it emits must always equal a fresh full recompute of the same query (Query.get()), which filters+sorts the whole collection and is the obviously-correct oracle.

The fuzzer

Each test drives one long random walk of create/update/delete operations over a small id/value space (so documents repeatedly cross the filter boundary, exercising the added/removed/modified transitions in _onBroadcast) and compares the latest emitted value against the oracle after every step. Covers sorted and unsorted queries across low/middle/high filter selectivity. A failing case replays from its seed + sorted flag + threshold.

The bug it found

When a document already in a query's result is deleted and recreated within the same broadcast batch, the two events coalesce in the event store into a single added event. The added handler only added documents passing the filter — it never evicted an already-cached document whose recreated value now fails the filter. So a stale snapshot lingered in the result:

doc 6 = 5  → passes filter, enters result
... later, same batch ...
delete(6); create(6, value 0)   → coalesces to `added`; value 0 fails the filter
→ before: result still shows the stale 6=5; after: 6 correctly removed

The modified handler already handled its analogous "was cached, now fails → evict" case; the added handler didn't. Fixed by evicting the cached entry there too. Confirmed the fuzzer fails against the unfixed code (e.g. seed 1001, round 34) and passes with the fix.

This is the second bug found by automated exploration in this hardening pass (after the PathRefStore ref leak in #35).

A note on test determinism

The fuzzer uses one long walk per test rather than many short trials: resetting the global store between trials schedules a broadcast that can race the next trial's observer (a test-isolation artifact, not a production issue). It also waits on a 1ms settle rather than a zero-duration one, so the broadcast's zero-duration timer and its microtask stream delivery are guaranteed to have completed before each comparison. With this shape it's deterministic — verified 0 failures across many thousands of rounds.

Test plan

• flutter test test/core/observable_query_equivalence_test.dart — 6 tests green; stable across 5+ repeated runs
• Verified the fuzzer fails against the unfixed _onBroadcast
• flutter test test/core — 112 tests green (run twice)

---

Generated by Claude Code

[Copilot]

Pull request overview

Adds a randomized equivalence test for ObservableQuery that compares incrementally-maintained results against fresh Query.get() recomputes, and fixes a stale-cache bug it surfaced: in the added/hydrated branch of _onBroadcast, an already-cached document whose recreated value no longer passes the filter was not evicted.

Changes:

• Evict cached doc in the added/hydrated branch when the new snapshot fails the filter, emitting a removed change snapshot symmetric to the modified branch.
• New fuzzer test driving long random create/update/delete walks against sorted and unsorted queries with varying selectivities.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
lib/observable_query.dart	Adds eviction + removed change snapshot when an added/hydrated event for an already-cached doc no longer passes the filter.
test/core/observable_query_equivalence_test.dart	New randomized equivalence fuzzer comparing emitted results against Query.get() across sorted/unsorted and multiple thresholds.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.