flowfile: the compiler knows every task output's declared type and the checker still says dyn — thread descriptors into env.Check

Observed behavior

Reference existence is validated — an unknown step or output is a positioned unresolved-reference diagnostic — but reference types are not: checkExpressionTypes declares every referenced name dyn (pkg/flowstate/v1/flowfile/celcheck.go:66-75), so an expression made of references is judged only for unknown functions, arity, and missing overloads. A type misuse across references (${steps.fetch.status_code.startsWith("2")} where status_code is an int) passes flow validate and fails at run time — on a durable run, potentially hours in.

The types exist at compile time and stop one step short of the checker: layer 3's whole point is that TaskDef descriptors are the single source validation, dispatch, completion, and generated docs already read (docs/ARCHITECTURE.md, "layer 3 is the single source of truth for capability" — docs/reference/tasks.md renders typed outputs from them today). Input declarations carry types (#177 slice 1), wait outputs are a known shape (timed_out, payload), and a loop's results/state are structurally known.

Desired outcome

env.Check sees declared types where they are knowable — task outputs from descriptors, declared input types, wait and loop shapes — and keeps dyn where dynamism is the honest answer (an http response's json, plugin tasks without descriptors). A type misuse across typed references becomes a positioned type-mismatch-class diagnostic in flow validate and the LSP alike, since both run the same check.

Acceptance criteria

• The int-vs-string case above is a compile-time diagnostic positioned at the expression, on validate and in-editor.
• The examples corpus stays green with zero new diagnostics — genuinely dynamic outputs produce no false positives.
• No second checker and no second environment: the existing checkExpressionTypes path grows declarations, per invariant 2.
• flow test's stub returns are checked against the same declared types, or the divergence is recorded (a stub that returns a shape the descriptor forbids is a test lying about production).

Adjacency

#177 (the full typed-workflows program — this slice needs none of its user-declared types, only the descriptors that already exist), #1301 (one bad expression suppressing other diagnostics — worth fixing before this raises the diagnostic count), #1377 (typed outputs on workflows would flow into the same declarations for call: results).

Start in pkg/flowstate/v1/flowfile/celcheck.go:66-75 at checkExpressionTypes and trace the existing env.Check path used by flow validate and the LSP. Read the layer 3 guidance in docs/ARCHITECTURE.md and typed output behavior in docs/reference/tasks.md, then inspect the examples corpus and flow test stubs. Done means typed references diagnose mismatches while dynamic outputs remain valid.