extproc: skip CONTINUE_AND_REPLACE when no body change is needed

[ChrisJBurns]

Description

Scope: this PR fixes the narrow case where the upstream cluster ext_proc
filter issued CONTINUE_AND_REPLACE even though nothing in the request
pipeline needed to mutate the body — no translation, no model override, no
retry, no backend HTTPBodyMutation. In that path, the upstream filter
replayed u.parent.originalRequestBodyRaw (snapshotted at the router phase),
which clobbered any body mutation applied by an earlier ext_proc filter in the
listener chain. After this PR, that path emits CONTINUE and leaves the body
alone.

The broader question of "preserve an intermediate filter's body mutation
across translation / retries / HTTPBodyMutation" is not addressed by
this PR. When wantBodyReplace == true, the replacement body is still derived
from the router-phase snapshot, so any intermediate filter's mutation is lost
on those branches. That's an architectural property of the snapshot-at-router-
phase design — the upstream filter has no view of Envoy's current body, by
design, to avoid re-buffering through ext_proc at the upstream phase. Fixing
it requires a separate design (either re-buffering at upstream, or
propagating a diff via dynamic metadata) and should be its own PR.

Root cause of the narrow bug. The existing safety net in
applyBodyMutation looks like it covers the no-config case:

if bodyMutator == nil {
    return bodyMutation
}

but it does not fire in practice. bodymutator.NewBodyMutator(nil, ...)
returns a non-nil struct even when its *filterapi.HTTPBodyMutation argument
is nil. So the early-return is dead code; the function falls through, calls
Mutate(originalRequestBodyRaw) on a no-config mutator (which returns its
input unchanged), wraps that in a BodyMutation, and the upstream filter
sends CONTINUE_AND_REPLACE with the original body — even though nothing
asked for a body replacement.

Fix:

• Add BodyMutator.HasMutations() predicate. NewBodyMutator's contract is
  unchanged (it still returns non-nil); callers can ask whether the mutator
  actually has anything to do.

• In upstreamProcessor.ProcessRequestHeaders, compute

  mutatorHasMutations := u.bodyMutator != nil && u.bodyMutator.HasMutations()
  wantBodyReplace := bodyMutation != nil || forceBodyMutation || mutatorHasMutations
  

  before calling applyBodyMutation. When false, skip the body-mutation work
  and emit CONTINUE instead of CONTINUE_AND_REPLACE, preserving translator
  header mutations (e.g. path rewrite), header-mutator output, and auth-handler
  header additions. Skip the content-length dynamic-metadata restamp on the
  CONTINUE branch since the body length did not change.

All existing CONTINUE_AND_REPLACE callers stay on the existing path (these
are the branches where this PR explicitly does not change behavior, including
the loss-of-earlier-mutation property documented above):

• Translator emits a body (AWS Bedrock/SigV4, OpenAI to Anthropic, OpenAI to
  Vertex, etc.) — bodyMutation != nil.
• Backend HTTPBodyMutation configured — HasMutations() == true.
• Retry — forceBodyMutation == true via u.onRetry().
• Streaming with IncludeUsage=false — forceBodyMutation == true.
• Model override — translator emits newBody via sjson.SetBytesOptions, so
  bodyMutation != nil.

Tests:

• Added Test_chatCompletionProcessorUpstreamFilter_ProcessRequestHeaders_BodyReplaceContract
  with two subtests:
  • no translator body, no mutator, no force -> CONTINUE exercises the
    previously untested path. Asserts Status == CONTINUE, no BodyMutation,
    preserved translator and auth-handler header mutations, and
    DynamicMetadata == nil (no content-length restamp).
  • translator body present -> CONTINUE_AND_REPLACE pins the existing
    behavior so a future refactor cannot quietly flip the contract back.

The coverage gap that let this ship: every existing t.Run("ok", ...) subtest
under Test_chatCompletionProcessorUpstreamFilter_ProcessRequestHeaders set
retBodyMutation on the mock translator, so the "translator returns nil body"
path was never exercised.

Related Issues/PRs (if applicable)

None.

Special notes for reviewers (if applicable)

• The fix is request-side only. Response-side code is untouched.
• NewBodyMutator's contract is intentionally not changed (it still returns
  non-nil even when given a nil *filterapi.HTTPBodyMutation). A predicate is
  added instead, to keep the blast radius local to ProcessRequestHeaders. If
  reviewers prefer changing the constructor to return nil in the no-config
  case, that is a larger refactor and a separate PR.
• The bug is in the generic upstreamProcessor[...], so the fix applies to
  every endpoint spec (chat completions, messages, embeddings, completions,
  GCP, AWS Bedrock, etc.). The new tests exercise one endpoint spec for the
  same reason the existing ProcessRequestHeaders table-test does — locking
  the generic method's contract.

[PatilHrushikesh]

if wantBodyReplace is true, it will still overwrite earlier mutation right?

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.73%. Comparing base (9a4b02c) to head (c8009c1).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2170      +/-   ##
==========================================
+ Coverage   84.72%   84.73%   +0.01%     
==========================================
  Files         144      144              
  Lines       21172    21189      +17     
==========================================
+ Hits        17938    17955      +17     
  Misses       2152     2152              
  Partials     1082     1082              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[ChrisJBurns]

if wantBodyReplace is true, it will still overwrite earlier mutation right?

Correct - when wantBodyReplace == true, the replacement body is still derived from u.parent.originalRequestBodyRaw (snapshotted at the router phase), so any mutation an earlier ext_proc filter applied between the router phase and the upstream phase is lost. That's an architectural limitation of the snapshot-at-router-phase design and not in scope for this PR. The narrower thing this PR fixes is the case where the upstream filter was issuing CONTINUE_AND_REPLACE even though nothing needed to mutate the body (no translation, no override, no retry, no HTTPBodyMutation) - that path used to always replay the original and clobber earlier filters; now it emits CONTINUE and leaves the body alone. I'll tighten the PR description to make that scope explicit.

[nacx]

Overall LGTM. @mathetake could you take a look as well?

[nacx]

/retest