V2's session runner was hardcoded to 25 steps regardless of the agent's
configured `steps` field, and never warned the model on its final
configured turn. V1 (`packages/opencode/src/session/prompt.ts:1314-1315,
1426`) does both — read `agent.steps ?? Infinity` per iteration and
append a synthetic assistant message from `max-steps.txt` on the last
step. This restores V1 parity so the planned V1→V2 cutover does not
regress user behavior.

## What changed

- `packages/core/src/session/runner/llm.ts`
  - Add `MAX_STEPS_PROMPT` constant — inlined verbatim copy of V1's
    `packages/opencode/src/session/prompt/max-steps.txt`.
  - At the top of `run()`, read the per-agent cap
    (`runAgent.info?.steps ?? MAX_STEPS`) once and use it as the for-loop
    bound. Previously hardcoded `MAX_STEPS = 25`.
  - Thread a `Ref<number>` (closure-scoped, reset on each `run()` call)
    so `runTurnAttempt` can read the current iteration index.
  - At the request build site in `runTurnAttempt`, read the index and
    re-read the agent's `steps` per call (to pick up mid-session swaps),
    then append the synthetic assistant message when
    `currentStep + 1 >= maxStepsForAgent`.
  - `StepLimitExceededError.limit` now reports the resolved cap (was
    always `25`).

- `packages/core/test/session-runner.test.ts`
  - 4 new tests in the existing `SessionRunnerLLM` describe block:
    1. `agent.steps = 3` raises `StepLimitExceededError({ limit: 3 })`
       (regression for anomalyco#30865).
    2. No `agent.steps` override → still defaults to 25.
    3. `agent.steps = 2` → 2nd (final) request's last message contains
       "CRITICAL - MAXIMUM STEPS REACHED" (regression for anomalyco#30866).
    4. `agent.steps = 5` → non-final turns do NOT carry the signal.

## Why

- The hardcoded cap ignored user-configured step limits (anomalyco#30865).
- Without the last-step signal, the model emits tool calls on the
  cutoff step that are killed mid-flight, wasting the final turn (anomalyco#30866).
- Both are V1↔V2 architecture gaps that block the planned V1→V2
  cutover. V1 already has the correct behavior; V2 now matches.

## V1-parity notes

- V1 reads `agent.steps ?? Infinity` per iteration; V2 reads once in
  `run()` for the for-loop bound and re-reads in `runTurnAttempt` for
  the `isLastStep` check. If the agent is swapped mid-session, the
  for-loop bound stays at the original cap (acceptable: it is a safety
  net; per-call cap detection is the actual enforcement).
- V1 reads `agent` from the last user message; V2 reads from the
  session. Minor known deviation; per-message agent switching in V2 is
  not yet supported.
- V2 retains its 25-step default when `agent.steps` is unset. V1 has no
  default cap (`?? Infinity`); V2's 25-step default is a strict
  improvement and is preserved.

## Verification

- `bun typecheck` in `packages/core` — clean
- `bun turbo typecheck` (23 packages) — all pass
- `bun test test/session-runner.test.ts` — 95 passed, 0 failed
  (4 new + 91 pre-existing)
- `bun test test/session-runner-model.test.ts
         test/session-runner-message.test.ts
         test/session-runner-tool-events.test.ts
         test/session-runner-tool-registry.test.ts
         test/session-runner-recorded.test.ts` — 37 passed, 0 failed
- `git diff --check` — clean

Closes anomalyco#30865
Closes anomalyco#30866