feat(opencode): add per-session budget limit

[HHrddtu]

Issue for this PR

No linked issue — new feature.

Type of change

• New feature

What does this PR do?

Adds an optional per-session budget that stops the assistant when the session cost reaches it, plus a small widget in the TUI sidebar Context panel to view and set it.

• budget field on the session schema (v1 + current), a nullable budget column on the session table, and a generated drizzle migration
• PATCH /session/{id} accepts budget (a number sets it, null clears it); sessions can be created with a budget
• Enforcement in the V1 prompt loop: after each step-finish the accumulated cost is compared against the budget; when crossed, a text part "Session budget reached. Increase the budget to continue." is appended and the loop stops. Re-prompting does not start a new provider turn until the budget is raised (guard at loop entry)
• TUI: the Context panel shows $0.42 / $2.00 budget (red when exceeded, unlimited budget when unset). Mouse scroll over the row adjusts it in $0.25 steps (scrolling below $0.25 clears it), clicking switches to an input — Enter saves, Esc cancels
• SDK types regenerated; the OpenAPI encoder drops the null arm of Schema.NullOr wrapped in Schema.optional, so the sdk build script now widens the update-side budget types to number | null

Known limitation: enforcement lives in the V1 prompt loop. The durable V2 runner does not compute cost yet (cost: 0 is hardcoded in Step.Ended), so budget stopping is a no-op there until V2 cost accounting lands.

Small extra: packages/core/script/migration.ts got a Windows fix — generatedMigrations split on / only, which broke migration generation on Windows. It was required to generate the migration for this PR.

How did you verify your code works?

• bun typecheck in packages/schema, core, opencode, tui, sdk/js — all clean
• bun test test/session in packages/opencode: 383 pass (new tests: budget stops the turn, exhausted budget exits the loop without an LLM request, budget round-trips through Session.Info)
• new packages/tui/test/feature-plugins/sidebar-context.test.tsx: headless widget tests (render, wheel up/down, clear, click-to-edit, Enter save, Esc cancel) using the OpenTUI test renderer with mock mouse/keys
• migration generator --check passes
• manually ran bun dev and exercised the widget with mouse and keyboard

Screenshots / recordings

Char frame from the headless test renderer (the Context panel of the session sidebar):

Context
0 tokens
0% used
$0.42 spent
$0.42 / $2.00 budget

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍