perf(viewer): deduce spaces lazily instead of the full database in SSR

[jamesdabbs]

Stacked on #254 (perf/viewer-ssr-isolate-cache) — this PR's base is that branch, so the diff here is only the deduction change. Merge #254 first.

What & why

The universal layout load runs during SSR, and refresh() ran the full deduction engine over every space on each request — then the client discarded it and recomputed the same thing after hydration. One logged invocation spent 440 ms of CPU rendering a single static reference page. Only 3 routes are prerendered, so every detail page (exactly what crawlers and deep links hit) paid this.

Gate the eager full-database deduction to the browser, and have checked(spaceId) deduce only the space it needs, on demand — so SSR of the trait page still renders the data it references. Side effect: removes the duplicate deduction passes the old test documented as a FIXME.

Instrumentation

Adds a deduce_space Workers Logs event (space id + derived count). After this change an SSR page render emits exactly one deduce_space (the requested space), and deduce_run (the full-database pass) disappears from runtime logs.

Tradeoff

SSR'd space pages show only asserted traits in the initial HTML; deduced ones fill in after hydration (unchanged from how the client already worked). If SEO on derived traits matters, prerendering all routes is the follow-up.

Validation

• pnpm --filter viewer tsc and validate (svelte-check) — clean
• pnpm --filter viewer test — 70 passed, 5 todo (the FIXME duplicate-pass assertions are gone)
• VITE_SITE=topology pnpm --filter viewer build

Manual deploy + telemetry check (topology worker pi-base-topology):

1. VITE_SITE=topology pnpm --filter viewer build
2. pnpm --filter viewer cf:deploy:topology (= wrangler deploy --env topology)
3. GET a non-prerendered /spaces/<id>/properties/<id> page.
4. In Workers Logs / observability:
   • Expect: exactly one deduce_space per render (the requested space), and no deduce_run from SSR invocations.
   • Compare the platform cpuTimeMs on the detail routes (request.routeId) before vs after: should drop from ~440 ms toward low tens of ms.

Reviewing

Easiest commit-by-commit; base is #254.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	pi-base-topology	e928786	Commit Preview URL Branch Preview URL	Jun 28 2026, 08:29 PM

[cloudflare-workers-and-pages]

Deploying topology with    Cloudflare Pages

Latest commit:	e928786
Status:	✅  Deploy successful!
Preview URL:	https://8c7ced7c.topology.pages.dev
Branch Preview URL:	https://perf-viewer-ssr-deduction.topology.pages.dev

View logs

[jamesdabbs]

Telemetry validation — deployed to pi-base-topology (version bc73d93f)

Deployed this branch (rebased onto main after #254 merged) to the topology worker and validated against live traffic + a controlled request batch.

✅ Eager full-database deduction eliminated from SSR

deduce_run events: 52 in the hour before deploy → 0 after. The full-database prover no longer runs during SSR.

✅ Lazy per-space deduction works, and SSR renders correctly

Raw SSR HTML for /spaces/S000001/properties/P000016 (server-rendered, pre-hydration) contains the deduced verdict and its proof:

Space S1 | Property P16 is Automatically deduced from the following: … Theorems T430, T431, ¬T782 ∧ T795, T650

So checked() deduces the one requested space on demand and the deduced trait + proof are in the initial HTML — not asserted-only. deduce_space fires ~once per SSR render (the instrumentation), and the response is compact (~9 KB). Valid pages return 200; nonexistent ones (e.g. /spaces/S000600/properties/P000200) correctly 404.

✅ Heavy-route CPU dropped sharply

$workers.cpuTimeMs on /spaces/[id]/properties/[propertyId]:

percentile	baseline (#254)	#251
p50	889 ms	265 ms (~3.4× lower)
p90	2621 ms	404 ms
p99	3654 ms	404 ms (~9× lower)

The residual ~265 ms is base SSR render + occasional cold-isolate bundle transform; the 0.6–3.6 s of per-request prover work is gone.

Caveats / honesty on the sample

• The post-deploy window is a low-traffic evening lull (~6–22 SSR invocations in 41 min), so the p90/p99 above rest on a small sample (p90≈p99 because there are only a handful of points). The median and the deduce_run→0 result are robust; the tail magnitude should be re-confirmed over a busier soak.
• That cuts in our favor: low traffic means a higher cold-isolate fraction (more cold parses), so a busy-traffic window should land at or below these numbers, not above.
• The baseline row is the perf(viewer): cache the parsed SSR bundle per isolate #254 afternoon busy window (13:00–16:00 UTC, thousands of samples).

Reproduce

• Query Builder: https://dash.cloudflare.com/78c505984bbdc3e69206eecb9471c4de/workers-and-pages/observability/investigate
• Filters: $metadata.service = pi-base-topology; evt = deduce_run / deduce_space; heavy-route CPU via $workers.event.request.path ~ ^/spaces/[^/]+/properties/[^/]+$ over $workers.cpuTimeMs.

Verdict: behaves correctly (deduced traits SSR'd with proofs), eager deduction eliminated, and per-render CPU down ~3.4× median / ~9× p99. Recommend one more soak through a busy-traffic window to firm up the tail percentiles before merge.