fix(coordinator): re-arm the self-telemetry interval on failure

[pfmos]

Problem

get_self_telemetry() sends the four-byte "self" form of CMD_SEND_TELEMETRY_REQ — opcode plus three reserved bytes, no pub key (meshcore_py/commands/device.py). Firmware answers it; MyMesh.cpp splits the command by frame length and has an explicit len == 4 self branch.

A non-firmware companion need not. An openHop virtual companion up to and including 1.1.1 requires the 36-byte contact form and rejects the short frame:

# openhop_core/companion/frame_server.py
if len(data) < 35:
    self._write_err(ERR_CODE_ILLEGAL_ARG)
    return

_last_self_telemetry_update advanced only in the success branch, so against such a node the interval gate never re-armed. The configured interval collapsed to the coordinator tick and the request was re-sent on every cycle, forever, each one logged at ERROR.

Measured on a live instance against an openHop companion, from the repeater's own frame log over 30 minutes:

Opcode	Command	Count	Cadence	Result
0x27 len=4	SEND_TELEMETRY_REQ (self)	357	~5.0 s	refused, every time
0x14 len=1	GET_BATT_AND_STORAGE	357	~5.0 s	ok
0x0a len=1	SYNC_NEXT_MESSAGE	37	~48 s	ok
0x38 len=2	GET_STATS	15	~2 min	ok

The configured interval was 300 s. Half of every companion frame the integration sent was a request that could not succeed, and it produced roughly 17k ERROR-tier log lines a day.

Change

Re-arm the interval on failure. The interval gates attempts, not successes, so the attempt is recorded before it is issued. A failure now waits the configured interval exactly as a success does. Behaviour against a node that answers normally is unchanged.

Level the log by whether a retry could ever help. Failures route through a new _log_self_telemetry_error, following the existing _log_get_msg_error pattern. An error code meaning the node can never satisfy the request (ERR_CODE_ILLEGAL_ARG, ERR_CODE_UNSUPPORTED_CMD) is reported once at WARNING, naming the cause and the remedy, then demoted to DEBUG. Every other failure keeps ERROR on every occurrence. A success clears the flag, so a genuine later regression is reported afresh.

Together: a node that cannot answer produces one actionable WARNING instead of ~17k ERRORs a day, and stops consuming a request slot every tick.

Note on the other side of this

openHop already fixed its half — 3513bab, "answer the self form of telemetry requests", 16 Jul 2026 — but it is on dev only, not in any tag, and dev is ~340 commits ahead of main. Until that ships, released openHop builds still reject the frame. This change is also worth having independently: it makes any peer that refuses any optional command degrade quietly instead of hammering it.

Tests

tests/test_self_telemetry_retry.py, following test_get_msg_error_log_level.py: the helper is AST-extracted from production source and exercised directly, since coordinator.py cannot be imported whole under the conftest stubs.

• unsupported code warns once, then demotes to DEBUG
• ERR_CODE_UNSUPPORTED_CMD treated the same way
• other payloads (including None) keep ERROR
• structural guard: no assignment to _last_self_telemetry_update may sit inside the try that issues the request — this fails on the pre-change source and passes after, so it pins the actual bug rather than the symptom
• a success must clear the reported flag

Verified locally: the 5 new tests pass, and the collectible suite is unchanged versus clean HEAD (104 passed vs 99, same 32 failures and 10 collection errors, all pre-existing TypeError: unsupported operand type(s) for | from PEP 604 syntax on the local Python 3.9 — CI runs a newer interpreter).