feat(LinuxTentacleE2E): Phase 12.M.L.C.3 — register Polling flavor E2E

Summary

Section C breadth: Polling flavor exercises a different code path than Listening, with operator-meaningful differences.

Aspect	Listening (C1.h)	Polling (this PR)
Mode resolution	no --comms-url	--comms-url set
Endpoint	/api/machines/register/tentacle-listening	/api/machines/register/tentacle-polling
Registrar	TentacleListeningRegistrar (bespoke HTTP)	TentaclePollingRegistrar wraps TentacleRegistrationClient
Payload field	uri (listening URL)	subscriptionId
Response validation	(none — gap caught by J.M.L.C.2)	EnsureRegistrationPayloadComplete

Why ship-blocking

Without this pin, regressions in polling-specific code ship silently:

• Mode resolution flips → --comms-url ignored, operator gets Listening flow when they wanted Polling
• Endpoint path drift (/tentacle-polling → /poll) → server returns 404
• subscriptionId field renamed → server can't correlate dispatches to the agent

Pairs with C1.h to pin BOTH flavor branches

• C1.h: Listening + /tentacle-listening + uri field
• C2.h (this PR): Polling + /tentacle-polling + subscriptionId field

Assertions

Assertion	Pins
exitCode == 0	Polling registrar succeeds with valid response
Stub path = /api/machines/register/tentacle-polling	Endpoint contract
X-API-KEY header propagation	Auth contract (parity with Listening)
Payload contains subscriptionId (Polling-specific)	Field-presence positive
Payload does NOT contain uri (Listening-specific)	Field-presence reverse-assert (no flavor confusion)
Config persisted with stub's ServerThumbprint	TLS pinning material round-trip

Fidelity tier

🟢 High (Rule 12.4): same fixture as C1.h; differs only in --comms-url arg + expected endpoint + payload field assertions.

Expected runtime: ~1-2s.

Test plan

• Linux E2E workflow runs (manual workflow_dispatch after merge)
• C2h_RegisterPolling_HitsPollingEndpointAndPersistsConfig passes within ~3s
• No regression on existing 35 Linux E2E tests

🤖 Generated with Claude Code