[DO NOT MERGE] POC: Tessar query materialization

[christse]

Background and Goal

Tessar is a DO NOT MERGE experiment for shared dashboard lists and statistics. Opted-in card definitions use existing Boxel queries and computeVia; the indexer persists query watches and computed outputs, while readers consume complete snapshots without reconstructing the input graph. Pending or failed work cannot appear as a current successful result.

The final guarded build passes 20 connected Chromium reads and nine mutations on 12,550 fabricated records: 800 ms median / 831 ms p95 interactive readiness, zero input-card/search requests, and at most 5,222 ms acknowledgement-to-display delay across those mutations. Worker interruptions before and after source publication recover within the fixed 10-second deadline; a new authenticated browser opened during pending publication also passes, without stale initial HTML. Seven focused browser tests / 80 assertions, six Postgres cases, four generator/browser-oracle tests, the sliding-sync regression, and modified-package lint/type checks pass.

This is an implementation checkpoint, not a completed POC. A preceding matched stable-data series observed 674 ms median for Tessar versus 13,576 ms for the corrected getCards reference, but the reference fails write freshness, so this is not a qualified overall speedup. The complete 10× HTML rebuild still has one Chrome render timeout after bounded prewarming fixed an earlier worker heap failure. JSON-path measurements are explicitly labeled; controlled 1× and larger matching-set runs, broader lifecycle coverage, and full application workflows remain. No additional AI-generation capacity is claimed.

Where to start

• docs/tessar-query-materialization-plan.md records scope and unfinished gates. docs/tessar-performance-report.md, its exportable chart, and sanitized results preserve the evidence and failed runs.
• packages/runtime-common/tessar-index-publication.ts connects reverse matching, concrete dependencies, durable dirty state and guarded publication to the existing index transaction. The first-watch registration race has a failing-before/passing-after regression.
• packages/runtime-common/tessar-materialization.ts, base snapshot support and the host store implement the versioned read contract. Initial HTML shares the readiness/revision checks; HTML-only rendering consumes published values, while indexing still computes.
• scripts/tessar/ contains deterministic generation, independent full-output oracles, real Chromium checks, concurrent reads, worker interruptions and benchmark tools. Presets are smoke, 1× (1,255 records) and 10× (12,550); no production records are inputs.

Key decisions and non-obvious mechanics

• The resolved Boxel query AST remains authoritative. Indexed routing terms select conservative candidates; the existing SQL compiler verifies old/new matches. Empty queries, membership exits, content edits and transitive dependencies participate in invalidation.
• Watch replacement, output publication and dirty markers share a pinned transaction; rendering stays outside it. Generation guards reject work racing the first watch or newer source publication. The existing queue handles worker recovery.
• The browser gate requires a live realm subscription and removal of server markup. First-sync and reconnect revalidation close the tested notification gaps. Historical DOM-only timings remain provisional. Anonymous published pages need a separate freshness contract.
• Initial HTML rejects queued/failed or older materializations, and realms with active materializations use no-store instead of deployment-ETag validation. Explicit drill-down can hydrate a selected source; ordinary snapshot display does not.
• Three private-fork application projections pass a fresh full JSON/HTML build and feeder-to-dashboard browser checks with fabricated data. Full capture/generation flows and personal overlays remain. No deployment realm or shared staging runtime was modified; this draft remains DO NOT MERGE with auto-merge disabled.

[github-actions]

Preview deployments

• Host staging preview

• Host production preview

Host Test Results

    1 files      1 suites   2h 12m 22s ⏱️
4 743 tests 4 729 ✅ 14 💤 0 ❌
4 758 runs  4 744 ✅ 14 💤 0 ❌

Results for commit fead7a1.

Realm Server Test Results

    1 files    210 suites   1h 14m 13s ⏱️
2 764 tests 2 762 ✅ 0 💤 2 ❌
2 803 runs  2 801 ✅ 0 💤 2 ❌

Results for commit fead7a1.

For more details on these errors, see this check.