[zhongkechen]

Issue #, if available

Fixes #861.

Description of changes

Invocation input construction cleared pending update IDs before delivery succeeded,
while a valid checkpoint-free resume did not acknowledge the input sequence. Retain
those IDs until a valid response, then consume only operation versions covered by
that input. Failed delivery retains updates for retry, newer changes stay pending,
and an invocation acknowledgment cannot retreat a newer checkpoint watermark.

Regression coverage uses wait completion, so this fix is independent of #860's
callback producer change. Public-handler tests cover memory/filesystem stores,
failure before SDK input delivery, legacy stored state, multiple resumes, and saved
steps executing once. No existing test expectations change.

The existing
test_pending_response_is_valid_when_an_operation_completed_after_the_handler_saw_it
fixture now supplies concrete operation/update collections read by the real
acknowledgment path; its assertions are unchanged.

Design

• Design note (rule 3): not needed; this repairs acknowledgment behavior using
  existing fields and methods, with no public API change.
• Simpler options considered (rule 2): moving the existing unconditional clear to
  invocation completion would consume newer changes. Reuse the existing
  per-operation versions and advance_handler_seen; no new delivery ledger is
  needed. Keep the watermark monotonic while pruning IDs covered by each actual
  acknowledgment, including equal/older acknowledgments.
• Root cause (rule 5): preparing input was treated as successful delivery, and
  successful invocation completion never acknowledged checkpoint-free input.
• Net change: +9 source lines; zero public API members added or removed.

Assumptions

• Deployed Lambda parity for the acknowledgment point is unverified in this
  PR; this implements the contract described in [Testing]: Acknowledge invocation input without losing pending updates #861/fix(testing): match replay state to the service #795.
• Legacy persisted update IDs without per-operation versions retain the existing
  sequence-zero interpretation. Tests load that legacy format and preserve newer
  versioned changes.

Decisions not covered by the issue or the design note

• Use wait completion as the regression producer so the fix validates independently
  of [Testing]: Record callback terminal completions as updated operations #860. A callback supplies only the later suspension barrier.
• Keep the explicit mark_state_delivered helper for existing callers; the executor
  now uses versioned acknowledgment instead.

Validation

• Python 3.13.17, Hatch 1.16.5.
• Final focused regression suite against baseline d7ae955: 26 failed, 10 passed;
  baseline import path verified before running. With this fix: 36 passed.
• hatch run dev-testing:typecheck: passed (137 source files).
• Package hatch fmt --check: passed.
• hatch run dev-testing:test -q --tb=short: 1,784 passed. Reran with loopback
  access after the sandbox blocked the HTTP test servers.
• hatch run dev-testing:js-examples: 123 test files passed, 150 tests passed,
  zero retries; four selected files have no cloud tests. JS SDK commit
  3deb8389531d35642e2d4657a69575522dc7a7ea.
• git diff --check: passed.

Checklist

• Every new branch has a test that fails without it (rule 4).
• The first commit adds regression tests; before/after evidence is above (rule 5).
• No existing test expectations changed (rule 5).

By submitting this pull request, I confirm that you can use, modify, copy, and
redistribute this contribution, under the terms of your choice.