test: cover callback handlers, fetch, metrics-only, payload capture in local integration harness

[joeyzhao2018]

What this is

PR 2b of the dd-trace migration roadmap (Phase 0). PR 2a brought the local RIE harness to
coverage parity with the old AWS suite; this PR goes beyond the old suite, adding the
first four customer usage modes that have zero end-to-end coverage on any surface — L1
specs cover the units, but nothing pins these paths through a real invocation.

It moves no library code. Fixtures, case registrations, docs, and goldens only — all
goldens captured from the current pre-migration implementation.

Base: this branches off joey/l2-coverage-parity (PR 2a) and should land after it.

The four cases

Case	What it pins	Why it matters
manual-callback	Callback-style (event, context, callback) handler under manual datadog() wrap, through the RIE invoke (async completion via setTimeout, so the wrapper can't mistake it for a sync return)	The migration spike broke exactly this seam: the new host's tracePromise wrapper replaced promisifiedHandler's call site, and callback handlers returned null to API Gateway. handler.spec.ts pins the units; until now nothing pinned it end to end
cjs-fetch-requests	Global fetch outbound in redirect mode; mock echo shows the injected x-datadog-* headers (18/18 requests carry all four)	On Node 18+ fetch is undici, instrumented by dd-trace's undici plugin — a different injection path than the http/https plugin that cjs-http-requests pins via axios. Main-path pattern for Node 18+ customers
manual-metrics-only	DD_TRACE_ENABLED=false: enhanced + custom metrics still flush (via DD_FLUSH_TO_LOG), and no aws.lambda span, no trace JSON, no dd.trace_id log correlation	The metrics-only customer configuration — real and non-rare — had no end-to-end coverage. Return-value goldens are byte-identical to manual-send-metrics (verified), so the toggle is pinned to change only the tracing surface
cjs-capture-payload	DD_CAPTURE_LAMBDA_PAYLOAD=true in redirect mode: function.request / function.response span tags holding the captured payloads	Documented feature with no golden; the config-wiring phase is exactly where payload capture could drift invisibly

Docs and ledger

• Harness README: case-table entries for the four cases, plus a "deliberately not
  covered locally" list in the emulation-gaps section — response streaming /
  time_to_first_byte, direct-API and KMS/Secrets metric paths, aws-sdk client spans,
  durable execution — each with its owning suite, so the scoping discussion doesn't recur.
• migration_parity.md: the callback and payload-capture rows now point at the new
  goldens; new dd-core ownership rows for fetch/undici injection and Lambda profiling
  (DD_PROFILING_ENABLED is a documented, shipped feature that had no ledger row — the
  capability lives in dd-trace's profiler, covered by its suites plus the L3 Profiling
  row).

Verification

All four cases green in strict compare mode on the node22 and node18 legs against shared
goldens — no per-runtime overrides needed. manual-callback additionally verified on the
node26 preview leg (newest RIC + callback style was the riskiest combination). The CI
workflow runs every case per runtime leg by default, so no workflow change is needed;
expect 16 cases × 5 runtimes = 80 legs after this lands on top of PR 2a.

Remaining PR 2b items (not in this PR): kinesis/eventbridge/sns-sqs event goldens,
negative-toggle variants (DD_ENHANCED_METRICS=false / DD_LOGS_INJECTION=false), and an
authorizer encode/decode fixture.

Trying it locally

RUNTIME_PARAM=22 CASE_PARAM=manual-callback ./integration_tests_local/run.sh
RUNTIME_PARAM=18 CASE_PARAM=cjs-fetch-requests ./integration_tests_local/run.sh
RUNTIME_PARAM=22 CASE_PARAM=manual-metrics-only ./integration_tests_local/run.sh
RUNTIME_PARAM=22 CASE_PARAM=cjs-capture-payload ./integration_tests_local/run.sh

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.