fix(automations): raise iteration cap to 50, surface update errors

[mgoldsborough]

Problem

A scheduled automation ("daily users" on a tenant) reported No output captured for this run with Stop: max_iterations. These are one phenomenon: the agent called a tool on every iteration up to the cap — including the final wrap-up step — so the loop exited via the iteration ceiling instead of the model saying "done." The run's output only accumulates assistant text blocks, so a run that ends mid-tool-call has empty output → the reader's "No output captured" fallback.

Two root contributors:

1. The automation iteration budget is too small. The create-form default is 5, and the hard cap is 15. Under progressive disclosure the agent spends ~2–4 iterations on tool discovery (nb__search → promote → call) before real work starts, so 5 leaves almost nothing for the actual task + synthesis. The run we debugged spent 2 of 5 steps on manage_workspaces/manage_tools discovery alone.
2. The UI swallowed the save error. Trying to raise maxIterations to a value above the cap was rejected server-side (maxIterations must be between 1 and 15), but handleUpdate caught it in a silent catch {} and refetched — so the field snapped back and it looked like "save isn't saving."

Changes

• Cap 15 → 50 (the engine's absolute MAX_ITERATIONS) in validateAutomationFields, the platform schema description, and the type doc comment.
• Create-form default 5 → 25 (DEFAULT_MAX_ITERATIONS) and input max 15 → 50.
• Surface update errors: handleUpdate lets the rejection propagate (still refreshes in finally); saveField renders it in the detail view's existing error banner.
• Test: server.test.ts now pins the 1..50 bound (rejects 51, accepts 50 and 25).

Rationale

The iteration cap is a runaway-loop backstop, not a cost budget — per-run cost is bounded directly by maxInputTokens (200k) and maxRunDurationMs (120s). A run that hits the cap has already spent the tokens and yields nothing, so a low cap converts "expensive but useful" into "expensive and useless." 50 matches the engine ceiling and the chat surface.

Verify

bun run verify green (static + unit + web + integration + smoke); build:bundles rebuilds the automations UI clean.