docs(cache-policy): verified provider healing profiles in CK inventory

Source-grounded ground-truth pass on the OpenCode<->Pi quirk divergences ahead of
the MC-under-SUBC migration. Verified every divergent quirk is still valid (none
obsolete) and added a 'Provider healing profiles' section establishing the
organizing principle: quirk_pass_work = provider_requirement - serializer_healing,
a per-path property.

Verified (current source + harness-owner confirmation, not stale comments):
- @ai-sdk/anthropic 4.0.1 (latest) groupIntoBlocks still merges consecutive
  assistants and pushes thinking in source order (no reposition); OpenCode 1.17.10
  ships 3.0.82. stripReasoningFromMergedAssistants stays required.
- OpenCode AI-SDK ProviderTransform.message empty-filter covers anthropic+bedrock
  only (exact model.api.npm, excludes google-vertex/anthropic); a separate V2 core
  runner path heals universally.
- Pi 0.80.2 serializers heal empties universally (drop empty text/thinking,
  auto-fill reasoning_content='') and never merge consecutive assistants.

Migration consequence: on the owned/llm-runner path the residual can be ZERO if
llm-runner heals like Pi - Pi is named the reference implementation and the
llm-runner column is the spec to fill.

Co-authored-by: Alfonso [Magic Context] <288211368+alfonso-magic-context@users.noreply.github.com>

Author: ualtinok

docs/cache-policy/ck-message-field-inventory.md

@@ -103,6 +103,59 @@ is fixed across defer passes, and a provider/model switch is a HARD bust, so the
 emitter cannot vary mid-epoch. That fixed-determinism is what makes CK-level
 byte-stability imply wire byte-stability uniformly (no per-path vector binding).
 
+## Provider healing profiles (verified ground truth)
+
+The organizing principle for every bucket-(b) row:
+
+```
+quirk_pass_work(provider, path) = provider_requirement - serializer_healing(provider, path)
+```
+
+A "quirk" is never `f(provider)` alone. The same provider needs different
+treatment on different paths, because each path's serializer heals a different
+slice of the provider's requirements. So the quirk pass on a path applies only
+the RESIDUAL the path's serializer leaves unhealed. This was verified against
+current source + the two harness owners (not from stale comments):
+
+**Genuine provider requirements** (these are real, not harness bugs):
+- Anthropic / Bedrock: no empty text/thinking content blocks on the wire;
+  thinking blocks in the latest assistant turn must be unmodified and not
+  repositioned after text; redacted thinking + signatures preserved verbatim.
+- OpenAI-compatible subset (Kimi / Moonshot / DeepSeek-V4-thinking): a tool-call
+  assistant message must carry a non-empty `reasoning_content` field.
+
+**Serializer healing coverage, by path (verified):**
+
+| Requirement | Pi serializer (0.80.2, tarball-verified) | OpenCode AI-SDK path (`ProviderTransform.message`, 1.17.10) | OpenCode V2 core runner (`to-llm-message.ts`) | llm-runner (owned path) |
+|---|---|---|---|---|
+| drop empty text | universal | **anthropic + bedrock only** (exact `model.api.npm`; excludes google-vertex/anthropic) | universal | **TBD - must specify** |
+| drop empty thinking | universal | anthropic + bedrock only | universal | **TBD** |
+| auto-fill `reasoning_content=""` | yes (`requiresReasoningContentOnAssistantMessages`) | not in this path | n/a (native lowering) | **TBD** |
+| no consecutive-assistant merge (avoids interleaved-thinking 400) | **never merges** - one provider message per assistant | **merges** (`@ai-sdk/anthropic` `groupIntoBlocks`, still unfixed in 4.0.1 latest; OC ships 3.0.82) | n/a | **should follow Pi: do not merge** |
+| preserve redacted thinking + signature | yes | yes | yes | **TBD** |
+
+**What this proves for the migration (the load-bearing point):**
+
+- MC's per-harness divergence today is exactly the residual: Pi heals
+  universally so MC clears reasoning to empty + drops signature for ALL
+  providers and omits the merge + structural-noise workarounds; OpenCode's
+  AI-SDK path heals anthropic/bedrock only so MC keeps the `[dropped]` fallback
+  for copilot/openai-compat (#135) and the `stripReasoningFromMergedAssistants`
+  workaround (because @ai-sdk still merges + interleaves thinking).
+- **On the owned/llm-runner path the residual can be ZERO.** If llm-runner's
+  serializer heals like Pi (drop empties universally, auto-fill
+  `reasoning_content`, do NOT merge consecutive assistants, preserve redacted +
+  signatures), then EmptyContent/Reasoning are pure llm-runner-serializer
+  concerns and MC's downstream quirk pass does NOTHING on the owned path. The
+  entire merged-assistant interleaved-thinking quirk class simply does not exist
+  if llm-runner never merges (Pi proves this is the clean design).
+- Therefore the llm-runner column above is the spec to fill: each "TBD" is a
+  serializer behavior llm-runner must implement to keep the owned-path quirk
+  pass empty. **Pi (`~/Work/OSS/pi-mono/packages/ai/src/providers/`) is the
+  reference implementation** - `anthropic.ts` empty-drop + one-param-per-assistant,
+  `openai-completions.ts` `reasoning_content` fill, `transform-messages.ts`
+  cross-model thinking handling.
+
 ## Where assembly happens: plugin gap-fills, owned/MITM assembles
 
 Family *assembly* (Anthropic pulls System-role messages to a top-level `system`