feat(blocks): codegen — emit each block once, per-block trace seam, multi-edge throw

[07prajwal2000]

Why

Codegen inlined each block's continuation as a source string, so a shared tail downstream of a diamond was re-emitted once per branch (multiplying artifact size, compile time, JIT work), and edgeTo silently dropped every edge past the first on a handle. This also blocked per-block tracing and orchestration backtracking, which both need a stable per-block boundary.

What

• Compiles each reachable block to a named async function ($block_N); a shared continuation is called, never duplicated. Reachability/emission order computed via one DFS pass (collectReachable), separate from emission.
• edgeTo throws when a handle has more than one outgoing edge instead of dropping the extras; validateEdges() surfaces this at compile time for the whole graph, not just the reachable subset.
• New per-block trace seam: ctx.trace?.recordSpan(...), gated by a $trace hoisted once per invocation — zero allocation/emission when unset. Terminal blocks report via a new node.complete() on EmitNode; response.ts migrated.
• Artifacts compile to a small factory (/* fluxify-compiled-route-factory */ marker) instantiated via instantiateCompiled; old inline-function artifacts still run so a rolling worker update doesn't break in-flight artifacts.

Bug found + fixed in review

Loop bodies (for/foreach, via EmitNode.body()) never marked their own block as "reported" before calling into the executor block, unlike next(). So when an executor block inside a loop threw, the loop block's own catch also fired — recording a second, misattributed failure span under the loop's blockId in addition to the executor's correct one. Fixed by setting $recorded = true before the body call, matching next(). Covered by a new regression test.

Test plan

• bun test packages/blocks — 118 pass, 0 fail (added: shared-tail dedup, linear-source-across-diamonds, loop executor precedence/response propagation, trace span emission, error-span attribution incl. the loop misattribution fix, multi-edge-throw)
• bun run --cwd packages/blocks lint (tsgo --noEmit) — clean
• Manually reverted the fix and reran the new regression test to confirm it fails without it

Closes #193