Telemetry page: both directions, every backend, and the counters behind them

[emir-hasanbegovic]

Description

The Debug Telemetry page was showing one direction, one backend, and one
misleading colour. GET /api/debug was eight inbound counters plus the
preferred backend's name, so on a Windows box with both drivers the page
only ever said "ViGEm"; it painted an idle-but-healthy host red because it
read backendAvailable (which means "the bus is open right now", not "the
driver works"); and it kept polling /api/debug twice a second for the life
of the tab once you had visited it.

GET /api/debug now carries rx, tx, audio and auth blocks plus the
host gauges the page had been guessing at, and the page is rebuilt around
them.

Server

Block	What it answers
rx	Accepted inbound messages by type, and the four ways a datagram is refused before any decoder runs (malformed = known opcode that failed its length guard, unknownType, runt, unknownToken)
tx	The direction that had no telemetry at all: datagrams and bytes, split across every message the server sends, plus unroutable / encryptFailed / oversize / sendFailed
audio	Mic frames accepted / too-late / dropped (a partition of every inbound frame), what reached the pad by decode, by Opus in-band FEC and by concealment, and outbound frames sent / suppressed as digital silence / encode-failed / lost to lock contention
auth	Client-API 401s split notPaired / badProof

Plus peakLoopUs, webPort, mdnsResponderActive, clientApiListening,
live connection and controller counts, and sessionsReaped.

Two bugs fixed on the way:

• sendto's return value was discarded in ClientAdapter, so an
  ENOBUFS or EHOSTUNREACH was invisible. It is checked now and feeds
  tx.sendFailed.
• maxLoopUs reads 0 nearly always. /api/debug zeroes it on read,
  but the receiver's thread-local high-water mark never resets, so the
  atomic only climbs again on a new all-time record. The window semantics
  are kept (the local bench harness reads them), and a new peakLoopUs
  folds each windowed read into the true peak.
• Also: webPort was read by the page but never sent by the server.

Hot path

Untouched, and the split is deliberate rather than incidental. Nothing was
added to the accepted-packet path:

• inbound counters sit only in the receiver's rejection branches and its
  cold non-gamepad dispatch branch;
• DispatchResult gained a handled flag so the receiver can tell a
  malformed frame from an unrecognised one without re-parsing, which is
  what keeps inner_dispatch.cpp free of globals (its test target does not
  link globals.cpp);
• no outbound message is reachable from the gamepad path, so counting in
  sendEncryptedPacket is free;
• audio counters run at 50 Hz per controller under locks those paths
  already hold.

Inbound byte counting is deliberately absent: it would cost an atomic add
per packet on the accepted path.

src/app/wire_stats.h is header-only with an inline global, so there is
no per-platform globals.cpp for a new counter to drift out of.

Page

Bidirectional flow diagram, six grouped sections (54 rows), a mirrored
in/out traffic chart, and a list of every backend the host reports with
its vendor, kernel/user mode, audio capability, lifecycle and
installed-versus-bundled driver version. Idle reads idle instead of error,
polling stops on navigation, and a pre-2.0.0 satellite degrades to em
dashes rather than rendering undefined. 80 debug strings across all six
locales.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change
• Refactor
• Documentation update

How Has This Been Tested?

• Unit tests (ctest / test_session_service / test_windows_platform)
• Manual testing on Windows
• Manual testing on macOS
• Manual testing on Linux
• Tested against a paired client end-to-end

Automated. Full ctest 35/35 on the MinGW lane, scripts/check-format.sh
clean against clang-format 22.1.4, scripts/check_core_purity.sh clean.

• New test_wire_stats (108 assertions): every opcode's mapping in both
  directions, the malformed-vs-unknown split, the deleted registration
  opcodes staying "unknown", the peak fold across a zeroing read, and
  reset() coverage checked generically through forEachCounter, so
  forgetting to reset a future counter fails the test rather than silently
  leaking across a rebind.
• test_receiver: handled asserted for every accepted message type and
  every length-guard rejection, including the subtle case that a
  well-formed frame the service then refuses on policy (no CAP_MIC) still
  counts as handled, because that is what separates rx.malformed from a
  policy drop.
• test_session_service: audio counters cross-checked against the existing
  concealment/reorder scenario (accepted, decoded, FEC-recovered and late
  all agree with submitMicAudioPcm call counts) and the
  silence-suppression scenario.
• test_status_json: golden shape updated, plus a new test pinning the
  counter blocks to /api/debug and keeping them off /api/status and the
  1 Hz SSE tick.

Page. Rendered headless (Chrome, DOM asserted, not just eyeballed)
against five payload shapes and three locales:

Scenario	What it proves
Busy Windows host, both drivers	Both directions populate; both backends named with versions
Idle host	The backend stage reads Idle, not error — the reported bug
ViGEmBus missing + helper missing	The flow does go red; every rejection counter surfaces
Live 1.1.0-dev server (the installed build)	New blocks degrade to em dashes, no NaN/undefined, both backends still listed
Linux/uinput payload	No Windows-only assumptions
?lang=de, ?lang=fr	Translations render, no raw keys, no overflow

Also verified: the rate arithmetic (including the counter-reset and
zero-elapsed guards) and that navigating away clears both timers, with the
showView wrappers in debug.js and logs.js chaining rather than
clobbering.

Not tested: an end-to-end run with a paired client streaming controller
audio, so the audio block's live values have not been observed against
real traffic (the counters themselves are unit-tested against real Opus).
macOS and Linux builds were not run locally.

Checklist

• My code follows the project's style guidelines (clang-format clean)
• I have performed a self-review of my code
• I have commented my code in hard-to-understand areas
• I have made corresponding changes to documentation
• My changes generate no new compiler warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes