revert(session): remove the empty-step guard that mis-flagged no-arg tool calls

[wqymi]

Summary

A pure revert of the empty-step guard (commit 0c73f7440, PR #1730, later narrowed by #1776). Nothing new is added — after this PR, packages/opencode is back to its pre-guard behavior on this path.

The guard classified a tool call whose input was {} as an "empty step" with NO PROGRESS. That is wrong: input: {} is completely legitimate for no-arg tools (list_apps, plan_exit, plan_enter, …) whose schema is z.object({}). Real evidence: a computer-use_list_apps call with input: {} that COMPLETED with real output was flagged as "NO PROGRESS", got a spurious nudge, and — after EMPTY_STEP_MAX_RECOVERY — hard-halted the turn with a bogus Empty tool call loop detected terminal error. A tool call that executed successfully IS progress.

The design cannot be narrowed into correctness: "the model called a tool with no arguments" is not distinguishable from "the model made progress" without per-tool schema knowledge the guard does not reliably have. So it is removed rather than patched again.

What was reverted

File	Change
src/session/prompt/empty-step-detection.ts	Deleted (99 lines)
src/session/prompt.ts	Removed the empty-step-detection import, handleEmptyStep + both branch call sites (main + fork), emptyStepStreak, hardHalt and its if (hardHalt) break in the outcome === "break" gate
src/flag/flag.ts	Removed MIMOCODE_EMPTY_STEP_MAX_RECOVERY
test/session/empty-step-detection.test.ts	Deleted
test/session/empty-step-guard-integration.test.ts	Deleted
test/session/invalid-output-continuation.test.ts	Removed the "repeated empty output … halts the turn" case + the now-unused Flag import

Net: 6 files changed, 520 deletions, 0 insertions.

Why the leaked-toolcall-marker detector was dropped

An earlier revision of this branch also added a second, unrelated mechanism: a leaked-toolcall-marker classify variant (matching a text part whose whole trimmed content is "call:", "code", or a bare invoked tool name), an autoRetryLeakedToolcallMarker retry ladder in prompt.ts, a LeakedToolcallMarkerError in message-v2.ts, the MIMOCODE_LEAKED_TOOLCALL_MARKER_RETRY_LIMIT flag, and their tests. All of it has been removed from this PR, for three reasons:

1. The defect no longer occurs. The "call:" preamble leak was a quirk of Claude Opus 4.8. That model is no longer in use (Opus 5 now), and the leak is not observed in normal operation.
2. It would be dead code. Shipping and maintaining a detector, a retry ladder, an error type, a flag and a test suite for a defect that no longer happens is pure carrying cost.
3. The detection itself carries risk. A legitimate one-word code text part emitted alongside a same-named tool would be a false positive that discards an entire otherwise-valid step. And negatively priming the model about call: in a system reminder plausibly makes it more likely to emit the marker, not less.

If the leak ever reappears on a model we actually ship, it should be addressed then, with evidence from that model — not pre-emptively here, coupled to an unrelated revert.

Note: the branch name still reads …-add-call-preamble-leak-retry-suppress for history/CI continuity, but the branch's scope is now revert only.

Verification

• bun typecheck — exit 0, 12/12 packages successful (no dangling references to any removed symbol).
• Affected suites all green: test/session/classify.test.ts + test/session/invalid-output-continuation.test.ts (37 pass / 0 fail), test/session/prompt.test.ts (11 pass / 0 fail), test/session/prompt-effect.test.ts (47 pass / 2 skip / 0 fail).
• grep -rniE 'leaked.toolcall|LeakedToolcallMarker|LEAKED_TOOLCALL_MARKER|leakedToolcallMarker' packages/ → 0 matches.
• grep -rniE 'isEmptyStep|handleEmptyStep|emptyStepStreak|EMPTY_STEP_|hardHalt|empty-step-detection' packages/opencode/ → 0 matches.
• Rebased onto latest origin/main (d0b4cc077).