Make the entry metadata core-private and delete the cursor-mismatch guards

[tobyhede]

Parent

#799. Design record: .scratch/render-execution-unit/spec.md.

What to build

Close the seam. With both callers on it, the entry metadata type stops being a parameter of anything and becomes private to core — a local value handed straight to the observation deriver. One producer, so nothing can disagree with it.

The observation deriver's two cursor-mismatch throws are deleted with it. They exist to catch a caller passing a cursor that disagrees with the snapshot; there is no caller-supplied cursor any more, so the bug class is gone rather than guarded.

Acceptance criteria

• The entry metadata type is not exported from core's barrel and is not a parameter of any exported function.
• Exactly one function constructs it.
• The cursor-mismatch throws are removed, and a test records that cursor agreement is now structural.
• Any test that hand-built an entry now goes through the real seam rather than a test-only escape hatch.
• pnpm run verify green.

Blocked by

• Move the re-entry disclosure onto the entry seam; collect emits a complete STEP_ENTERED #820

[tobyhede]

Closed by #827 (merged as c1bbfd4bb). Each criterion verified against merged main, not against the PR description:

• Not on the barrel, and not a parameter of any exported function. StepEntryMetadata occurs in packages/core/src only in runbook/execution-unit-entry.ts and events/execution-observation.ts; the two functions taking it (shouldProjectInlineLaunchIntent, withInlineLaunchIntent) are module-private. Zero occurrences in packages/{cli,mcp,claude-code-plugin}/src. packages/core/__tests__/events/entry-seam-barrel.typecheck.ts pins the barrel absence with @ts-expect-error and carries a positive control (ExecutionUnitEntry), so deleting the whole export block cannot satisfy it.
• Exactly one constructor. One : StepEntryMetadata = { in the tree, at execution-unit-entry.ts:410.
• Cursor-mismatch throws removed, and agreement recorded as structural. Gone from deriveStepEnteredEffect. execution-observation.test.ts:262 records why the mismatch is now unrepresentable rather than merely untested, and the unreachability that argument rests on is what the typecheck file asserts.
• No test hand-builds an entry. The remaining StepEntryMetadata references under __tests__ are prose and the deliberate @ts-expect-error unreachability assertions.
• pnpm run verify green locally; 41 pass / 0 fail / 2 skipping on the merge commit.