Closes #346.

Adds a "Reporter / log events" section to `CODE_STYLE_GUIDE.md`
covering the convention for porting reporter emissions:

- **Finding the upstream emit** — three call shapes to grep for in
  `pnpm/pnpm` (`globalLogger.<channel>.<level>(...)`,
  `logger.<level>({ name: 'pnpm:<channel>', ... })`, and direct
  `streamParser.write(...)` calls).
- **Channel mapping** — `crates/reporter/src/lib.rs`'s `LogEvent`
  enum is the canonical list, with each variant pinning
  `#[serde(rename = "pnpm:<channel>")]`. Includes the convention
  for adding a new channel (mirror the upstream TS shape
  field-for-field, add a wire-shape unit test).
- **Threading** — `<R: Reporter>` generic on every fn that emits,
  turbofish at the production entry point, install-scoped state
  via parameters where dedup matters (e.g. `link_file`'s
  `&AtomicU8`).
- **Where to put the emit** — match the upstream call site's
  position relative to side effects; cite the upstream permalink
  at a pinned SHA in the code comment.
- **Testing** — recording-fake DI per #339, a `static
  Mutex<Vec<LogEvent>>` per test, sequence assertion via
  `matches!`. Verify the test catches the regression.
- **What not to do** — don't reformat upstream messages, don't
  invent channels, don't change emit granularity in either
  direction.
- **Worked example** — `pnpm:summary` in `Install::run` (landed in
  #355): shows the upstream permalink, the pacquet emit, and the
  test that pins the sequence.

Adds a sixth bullet to `AGENTS.md`'s cardinal rule pointing at the
new section, so the "match pnpm" mandate explicitly covers log
emissions.

A style guide should let a porter read it linearly without
context-switching to GitHub. The four issue references (#345 for
the reporter origin, #347 for the channel-backfill audit, #339 for
the recording-fake pattern, #355 for the worked example's PR) were
project history, not convention — none of them carried information
the doc itself didn't already restate inline.

Replace each with the concrete fact the reader needs: the reporter
lives in `crates/reporter` (with its trait, enum, and sinks named
in place); to add a new channel, do X, Y, Z (instead of "the work
tracked under #347"); the recording-fake pattern is the unit-struct
+ static-mutex shape (described directly, no "from #339"); and the
worked example is just the code, no "added in #355".

Two nitpicks on #374:

- The "concurrent tests don't race on the mutex" sentence
  conflated two distinct properties (cross-test isolation, which
  comes from the static being scoped to a single test fn, vs
  cross-retry contamination, which is what the reset addresses).
  Reword to call those out separately.
- The "mirrors `lodash.throttle(...)` in upstream's
  `remoteTarballFetcher.ts`" claim cited upstream code without a
  pinned permalink, breaking the same convention the guide states
  for upstream citations elsewhere. Add the pinned link to
  `remoteTarballFetcher.ts:143-144` at SHA `086c5e91e8`.

…ull set

Per Copilot review on #374: the doc said "the canonical list of
channels lives in `LogEvent`", but `LogEvent` only contains the
variants pacquet currently emits — it's not pnpm's complete channel
set. The doc comment on `LogEvent` itself acknowledges this:
"Variants are added as pacquet starts emitting them."

Reword to: "the channels pacquet currently emits live in `LogEvent`",
with an explicit note that the enum is *not* exhaustive and that
porting a function whose upstream emits a channel `LogEvent` doesn't
yet have is the trigger for adding a new variant. Keeps the existing
"To add a new channel" instructions intact below.