refactor(translation): normalize Anthropic requests through IR

[nachiketb-nvidia]

What

• Normalize outbound Anthropic requests through the typed LlmRequest IR.
• Remove backend-owned JSON rewriting for system roles, reasoning, tool IDs, and thinking blocks.
• Scope provider extensions to their decoded source format so Anthropic-native fields do not leak across formats.
• Preserve exact native requests only when nested cache_control annotations require it.

Why

Request semantics should be normalized once by source decoding and target encoding. The previous implementation moved a large serde_json::Value post-processor into the translation crate, but still duplicated behavior already represented by the IR and made provider ownership unclear.

How

• Record the decoded source_format on LlmRequest.
• Decode Anthropic message-level system/developer turns into typed instructions.
• Decode OpenAI-style reasoning_effort and encode it as Anthropic thinking plus output_config.
• Re-emit top-level extensions only when their source format is Anthropic; retain native context_management and drop cross-provider extensions.
• Sanitize and collision-disambiguate tool IDs while encoding typed tool calls/results.
• Omit unsigned thinking and messages left empty after typed encoding.
• Keep the translation crate exact-round-trip contract unchanged; the native backend requests canonical encoding except for nested Anthropic cache markers that the current IR cannot represent.

What to review

• LlmRequest::source_format and provider-extension ownership.
• Anthropic decode/encode behavior for instructions, reasoning, tools, and thinking.
• The narrow exact-replay exception for native cache_control blocks.
• Removal of the duplicate backend normalizer.

Validation

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo test --workspace
• cargo test -p switchyard-translation
• cargo test -p switchyard-components --test adversarial_native_backends

[coderabbitai]

Walkthrough

Anthropic outbound requests now pass through shared translation and buffered normalization. The normalization handles endpoint-compatible fields, system content, tool-use IDs, and unsigned thinking blocks. Translation and native-backend tests verify the resulting request shape and replay behavior.

Changes

Anthropic outbound request flow

Layer / File(s)	Summary
Buffered outbound normalization crates/switchyard-translation/src/codecs/anthropic/buffered.rs	Preserved and newly encoded Anthropic requests are normalized by removing incompatible fields, promoting system/developer content, sanitizing tool-use IDs, and filtering unsigned thinking blocks.
Native backend translation routing crates/switchyard-components/src/backends/anthropic.rs	The backend always uses shared request translation, then applies the model override and extra body fields; local normalization helpers were removed.
Normalization behavior validation crates/switchyard-translation/tests/request_translation.rs, crates/switchyard-components/tests/adversarial_native_backends.rs	Tests verify field normalization, extension preservation, tool ID sanitization, thinking-block filtering, and replay message counts.

Estimated code review effort: 4 (Complex) | ~45 minutes

Poem

I’m a rabbit with packets to pack,
Anthropic requests hop on the track.
Thinking blocks shed, tool IDs grow neat,
Shared translators make flows complete.
Tests thump their paws: “The shape is right!”

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title is concise and accurately reflects the main change: moving Anthropic request normalization into the translation layer.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/switchyard-translation/src/codecs/anthropic/buffered.rs`:
- Around line 488-497: Update normalize_outbound_request and its callers to
accept mutable diagnostics and policy, using push_lossy to report every dropped
message or non-text block. In
crates/switchyard-translation/src/codecs/anthropic/buffered.rs lines 488-497,
report removed messages and avoid emitting an empty messages array; in lines
375-386, report non-text blocks discarded when lifting system/developer turns.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 495958e0-d442-4cd2-a1d4-0fb611cdef29

📥 Commits

Reviewing files that changed from the base of the PR and between 36b9636 and db602bc.

📒 Files selected for processing (4)

• crates/switchyard-components/src/backends/anthropic.rs
• crates/switchyard-components/tests/adversarial_native_backends.rs
• crates/switchyard-translation/src/codecs/anthropic/buffered.rs
• crates/switchyard-translation/tests/request_translation.rs

[coderabbitai]

🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win

Lossy normalization has no diagnostics path. normalize_outbound_request takes and returns a bare Value, so every drop it performs is invisible to callers even though the rest of this codec reports losses via push_lossy. Threading &mut Vec<Diagnostic> + policy through would cover both sites and give a place to guard the all-messages-removed case.

• crates/switchyard-translation/src/codecs/anthropic/buffered.rs#L488-L497: report dropped messages and avoid emitting an empty messages array.
• crates/switchyard-translation/src/codecs/anthropic/buffered.rs#L375-L386: report non-text blocks discarded when a system/developer turn is lifted.

📍 Affects 1 file

• crates/switchyard-translation/src/codecs/anthropic/buffered.rs#L488-L497 (this comment)
• crates/switchyard-translation/src/codecs/anthropic/buffered.rs#L375-L386

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/switchyard-translation/src/codecs/anthropic/buffered.rs` around lines
488 - 497, Update normalize_outbound_request and its callers to accept mutable
diagnostics and policy, using push_lossy to report every dropped message or
non-text block. In
crates/switchyard-translation/src/codecs/anthropic/buffered.rs lines 488-497,
report removed messages and avoid emitting an empty messages array; in lines
375-386, report non-text blocks discarded when lifting system/developer turns.