feat(lint): ban try/catch and non-null assertions, clean packages/ and lib/

[Makisuo]

What

Two new lint bans, plus the burndown that makes them land green in shared code.

• maple/no-try-catch — a new rule in scripts/oxlint-plugins/maple.mjs, firing on TryStatement. oxlint has no no-restricted-syntax, so banning a statement kind takes a plugin rule. The message names the primitive per case: Effect.try/tryPromise, Schema.fromJsonString, Schema.decodeUnknown{Effect,Option,Sync}, Effect.catch/catchTag/catchDefect, Effect.ensuring for a finally.
• typescript/no-non-null-assertion — the oxlint built-in, no plugin needed.

packages/* and lib/* source is now at zero for both: 38 try/catch blocks and ~110 assertions converted.

Why

A thrown exception is invisible to the type system, so it escapes the typed error channel the codebase relies on, and one catch block flattens every failure into a single branch — discarding exactly the distinction the tagged-error convention exists to preserve.

Scope, and why it is staged

I measured before wiring anything up. Repo-wide there are ~660 try blocks and 2060 !, of which 1429 assertions are in tests. Both rules are "error" with three overrides:

1. Tests off for both. A test asserts on a fixture it just built, so ! there documents the fixture; a test driving a throwing boundary needs catch to observe it.
2. packages/browser, packages/browser-session, packages/clickhouse-cli — no-try-catch off. These carry no Effect dependency at all; the two browser SDKs ship to customers under a bundle budget, so try/catch is the only error handling available to them. Worth knowing before anyone widens the rule: check package.json first.
3. Burndown list — apps/**, scripts/**, **/scripts/**, examples/**, alchemy.run.ts off for both. ~460 try/catch and ~500 assertions remain there. Leaving them on would land bun run lint with ~960 errors; the comment in the config says to shrink the list, not widen it.

How the fixes were done

Rather than 38 inline conversions, the repeated shapes got a seam:

• packages/ui/src/lib/try-sync.ts and packages/effect-sdk/src/shared/try-sync.ts — Effect.try + Effect.option behind an Option, for host calls that throw instead of returning (JSON.parse, localStorage, execCommand, Intl, BigInt, decodeURIComponent).
• packages/ui/src/lib/local-storage.ts — the read and the write of a preference behind one guard.
• guardFlush in packages/effect-sdk/src/shared/flush-core.ts — replaces the identical never-reject wrapper the client, server, and Cloudflare presets each kept a copy of.
• Parsing moved to Schema.fromJsonString / Schema.URLFromString / Schema.decodeUnknownOption wherever a schema was the honest answer.

Almost every prod assertion turned out to be a noUncheckedIndexedAccess artifact (arr[i]!, map.get(k)!, match[1]!) rather than an unsafe cast, so the fixes make the invariant real instead of adding Option plumbing:

• Drain held two mutually-exclusive nullable stores; now one union field, which also let a test drop a cast into internals in favour of a public clusters().
• pipe-dispatch's int(key, def?) returned number | undefined for every defaulted call — overloading it removed six assertions with zero call-site edits.
• soleValue() in query-helpers.ts for the eight length === 1 ? values[0]! : … branches; childAt() for the nine has-then-get pairs in the prefix-tree walk.
• The LRU's entryAt() names a broken slot (LruSlotError) instead of surfacing undefined two frames later.

Two latent bugs closed on the way

• traces-shared.ts: an attribute filter over an empty key list built its index prefilter from keys[0]!, emitting has(…, NULL). It now falls back to the exact predicate.
• execution/executor.ts: the capability-aware compile passed capabilities! through; it now falls back to baselineWarehouseCapabilities(), which generates the widest SQL.

Verification

• bun run lint clean (both .oxlintrc.json and the type-aware .oxlintrc.effect.json).
• bun run typecheck — 39/39.
• Every touched package's suite green: ui 638, domain 622, query-engine 1352, effect-sdk 122, widgets 186, query-engine-integrations 314, clickhouse-builder 244.

Note for the reviewer

packages/domain's project-manifest.test.ts > keeps a stable project revision for unchanged inputs flakes roughly 1 run in 3 under the full suite and passes in isolation. It is not in this change's import graph and reproduces without it — pre-existing, and worth a separate look.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}