…f istanbul

The integration coverage harness instrumented dd-trace's source to count lines:
nyc's require-hook in every spawned child, plus a pack-time istanbul rewrite of
the sandbox tarball. That transform is the dominant cost of a coverage run
(~4x wall time on spawn-heavy suites, enough to time out child_process tests on
slower runners) and it breaks source-map resolution, which is why several
fixtures opt out entirely.

Switch collection to native V8 coverage (c8). Every process in the tree writes a
raw V8 profile into one shared collector directory via NODE_V8_COVERAGE; after
mocha exits, merge-lcov converts each profile through v8-to-istanbul into the
same coverage/node-<version> lcov the upload step already reads. No source is
transformed, so ESM and source-mapped files are covered without a hook.

1. patch-child-process.js now injects NODE_V8_COVERAGE (plus a minimal bootstrap
   require that only re-installs the patch) into children, including those that
   replace env; the opt-out blanks NODE_V8_COVERAGE rather than omitting it,
   because Node copies the parent's value to a child otherwise.
2. child-bootstrap.js drops nyc.wrap()/the pre-instrumented writer and instead
   flushes via v8.takeCoverage() on SIGTERM/SIGINT, so a long-running fixture the
   harness stops with a signal still contributes coverage.
3. The in-process test:*:ci suites move from nyc to c8 via scripts/c8-ci.js.
4. scripts/patch-v8-to-istanbul.js corrects V8's line-coverage over-report on
   multi-line statements: V8 leaves an indented, un-taken ternary/logical arm
   reported as covered, so the patch zeroes a line covered only from its first
   non-whitespace column. This keeps c8's line numbers aligned with the
   statement-granular istanbul baseline Codecov was built against.

Deletes the pack-time instrumenter, the pre-instrumented writer, the sandbox nyc
config, and the per-sandbox finalize step — all of which existed only to get
istanbul counters into child processes.

The codeowners audit only scoped `**/*.spec.js`, `benchmark/sirun/**`, and the
agent config dirs, so every non-spec file under integration-tests/ — fixtures,
helpers, forwarder scripts — could land unowned without tripping the lint gate.
The new coverage-merge-lcov.spec.js was the first to surface it.

Add an `integration-tests/**` glob to the audit and assign every path it flags:
the appsec graphql and standalone-asm fixtures to asm-js, the pino fixture to
apm-idm-js, the log_injection / remote_config fixtures and telemetry forwarder
to apm-sdk-capabilities-js, and the shared helpers, init, memory-leak,
package-guardrails, startup and tsconfig entries to lang-platform-js. The
package-guardrails directory entry replaces the file-by-file flush.js rule the
audit flagged as fragile.

`NODE_V8_COVERAGE` is not private to the harness — Node's own test runner
(`node --test --experimental-test-coverage`), c8 and other tools read it from
the child env. Overwriting it unconditionally redirected such a fixture's own
coverage into our collector, usually leaving its assertions green against empty
data. When a child already carries a directory, leave it in place and record the
collector in a copy-back marker instead; the per-spawn opt-out likewise blanks
`NODE_V8_COVERAGE` only when it is the value the harness injected, preserving a
directory the child set for itself.

A forked child's own profile is folded into the collector on its `exit`, after
V8's single teardown write, so `v8.takeCoverage()` is never called in a
gracefully-exiting child — calling it splits the child's execution counters
(each call resets them), which would corrupt the very coverage the child's own
tooling reads. The forceful stop paths still flush in-child, where the reset is
moot because the process is being terminated.

Reintroduce the Windows IPC flush sentinel in `stopProc`/`child-bootstrap`:
Windows SIGTERM is forceful and skips the signal hook, so a connected child is
asked to flush and exit cleanly on the sentinel, and its preserved-directory
profiles are copied on the ensuing exit.