js/deltic: recognize deltic values by protocol brands, not class identity

[lannbot]

Implements #48: the three sites that recognized deltic values by class identity now use the @deltic/protocol recognition predicates (brand symbols in the Symbol.for registry), which the pinned release (0.1.0-pre.g50778a1) already re-exports from @deltic/runtime/embedder — no pin bump.

The migration

• js/deltic/websocket.ts sendViaStream error wrap: instanceof WitError → isWitError. This was the real exposure: in a multi-copy graph, a WitError minted by a different runtime copy failed instanceof and was flattened to { tag: "other", val: String(error) } instead of passing its structured payload through.
• js/deltic/websocket.ts collectByteStream: instanceof Stream → hasBrand(stream, STREAM). Milder than the issue implies: deltic's Stream is async-iterable, so a foreign-copy handle previously fell into the for await fallback and was still consumed correctly — misrecognition cost the batched-read fast path, not correctness. Now it dispatches to the intended branch.
• examples/deltic-demo/run.ts error display: instanceof WitError → isWitError.

WitError stays imported for construction (witError(), the SendViaStreamError wrap); only recognition migrates.

Tests

Two new unit tests hand-roll branded values from nothing but the literal registry keys (deltic.witError/1, deltic.stream/1) — per the protocol contract a hand-rolled brand carrier is a legal value from any copy, so the tests prove recognition without class identity, which was untestable under instanceof short of loading two real runtime copies. Both fail against the pre-migration code (verified by mutation). The literal /1 keys are deliberate: an upstream brand-generation bump fails here and forces the recognition sites to be revisited.

MODULE-IDENTITY comments

As the issue anticipated, the pin-discipline comments (deno.json ×2, READMEs ×2, AGENTS.md, run.ts, worker-entry.ts, deltic-carrier.ts, demo header) narrow to the gate's remaining rationale: one runtime/translator plan-format pairing per graph, the runtime's refusal of cross-copy stateful handles, and the single-version review surface. The exam-deltic gate itself is unchanged.

One nit surfaced while implementing: the issue cites amendment A8; the pinned sources label the brand vocabulary A9 (renumbered upstream), so the comments cite the contract section by name instead of by amendment number.

Gates

• just deltic-module-check — 14/14 (incl. the 2 new tests)
• just conformance-ct — 4 targets, 220 results, 0 failing; matrix.md unchanged
• just exam-deltic — pin gate green
• deno check of examples/deltic-demo/run.ts under the driver config (not covered by any existing gate)

Fixes #48

[lannbot]

Resolved the conflicts by merging main: everything re-applied under the A10 rename (isWitError → isComponentException, payloads {tag, val} → {kind, value}) and the caret-honest 0.1.0 pin regime — the narrowed MODULE-IDENTITY comments now carry main's resolved-version/caret wording.

One fact worth recording: the A10 rename did not change the brand vocabulary — COMPONENT_EXCEPTION is still Symbol.for("deltic.witError/1") at protocol 0.2.0 — so the hand-rolled-brand tests' literal keys are unchanged, and the test comment now notes that key spellings are stable across upstream renames.

Gates re-run post-merge: deltic-module-check 14/14, conformance-ct 4 targets / 220 results / 0 failing (matrix.md unchanged), exam-deltic (now reporting the resolved 0.1.0), and deno check of the demo runner.