The dropped-block diagnostic is computed, carried through the adapter result, and never persisted — no artifact carries the per-kind counts #697

PR #660 (#625) added the per-kind dropped-block counter so an operator can
reconcile a thinking-enabled request's usage against the blocks that survived
translation. At abd1dcf4 the counter is produced and carried on the in-memory
result and read by nothing outside the test suite — no step report, no scan
aggregate, no results field, no Go consumer.

Mechanism (at abd1dcf4)

• libs/openant-core/utilities/llm/adapter.py:226 — the field:
  dropped_block_kinds: dict | None = field(default=None, repr=False)
• libs/openant-core/utilities/llm/providers/anthropic.py:378 — init {}
• libs/openant-core/utilities/llm/providers/anthropic.py:401-402 — increment
  per dropped block kind
• libs/openant-core/utilities/llm/providers/anthropic.py:474 — passed onto
  the CompletionResult

All five sites are producer-side. Nothing consumes the value.

Executed repro (re-run at abd1dcf4)

$ git grep dropped_block_kinds abd1dcf4 -- libs/openant-core ':!libs/openant-core/tests/'
abd1dcf4:libs/openant-core/utilities/llm/adapter.py:226:    dropped_block_kinds: dict | None = field(default=None, repr=False)
abd1dcf4:libs/openant-core/utilities/llm/providers/anthropic.py:378:    dropped_block_kinds: dict[str, int] = {}
abd1dcf4:libs/openant-core/utilities/llm/providers/anthropic.py:401:            dropped_block_kinds[str(kind)] = (
abd1dcf4:libs/openant-core/utilities/llm/providers/anthropic.py:402:                dropped_block_kinds.get(str(kind), 0) + 1)
abd1dcf4:libs/openant-core/utilities/llm/providers/anthropic.py:474:        dropped_block_kinds=dropped_block_kinds or None,

$ git grep -n "\.dropped_block_kinds" abd1dcf4
abd1dcf4:libs/openant-core/tests/test_issue625_thinking_control.py:23:  ``CompletionResult.dropped_block_kinds`` (a COUNT, never a token split —
abd1dcf4:libs/openant-core/tests/test_issue625_thinking_control.py:175:    assert result.dropped_block_kinds == {"thinking": 2, "redacted_thinking": 1}
abd1dcf4:libs/openant-core/tests/test_issue625_thinking_control.py:181:    assert result.dropped_block_kinds is None

$ git grep -n "dropped_block\|DroppedBlock" abd1dcf4 -- apps | wc -l
0

5 producer-side hits in production, 3 reader hits (all in one test file), 0 Go
consumers. #660's body claim that the counter "gives the reconciliation" is
broken at HEAD: the number never reaches an operator.

Falsifier

"any step report / scan aggregate / results field carrying the per-kind dropped-block counts."

One-line sibling (same surface, not a separate issue)

The declared SDK floor anthropic[bedrock]>=0.40.0
(libs/openant-core/pyproject.toml:11) predates the thinking kwarg — absent
at 0.40.0/0.46.0, present at 0.47.0 (wheel diff: 0 vs 20 occurrences).
libs/openant-core/requirements.txt:2 pins anthropic[bedrock]==1.7.0, so
there is no reachable consumer today; the floor is only wrong for an
installer that resolves the declared minimum.

Dedup anchors

Searched the live issue set for dropped_block_kinds (0 hits),
accounting_errors (0 hits), cost_incomplete, _summary.json.

• No request-side thinking control, and the response translation discards unsupported thinking blocks while forwarding their usage (an enhancement: expose the policy, keep the default) #625 (closed) asked for the diagnostic — delivered by feat(llm): request-side thinking control, effective-policy records, dropped-block diagnostic (#625) #660. This issue is
  about the persistence, not the design.
• registry: cache multipliers exist only for the anthropic family — openai/google/openrouter/bedrock cached usage reads cost_incomplete (bedrock under-reports) #690 / Registry census follow-up: the 4.5-generation Claude families (opus-4-5, sonnet-4-5 + dated forms) have no record under any provider; the claude-opus-4-6 status looks stale #648 are the registry/cache-pricing surface; no overlap with the
  dropped-block counter.
• No zero-reader duplicate exists.

Fix direction

Thread the counter into the same channel the other per-call diagnostics use:
the step report's token_usage block (libs/openant-core/core/step_report.py)
and from there the scan aggregate — present-only, so a healthy run's artifacts
stay byte-identical, matching the #605 counter's own convention.