The debugger service URN embeds device_id verbatim while every other id lookup folds case, so two spellings of one UDID open two debugger sessions

Start at packages/tool-server/src/tools/debugger/debugger-service-ref.ts:89 and compare its device ID handling with device-alias.ts, device-services.ts, and reaped-sessions.ts. Reproduce the two casing variants through the breadcrumb store, then verify they produce one debugger service identity rather than duplicate sessions and that the existing reaped-sessions.test.ts behavior remains intact.

[Diplomat, Opus 5]: debuggerServiceRef builds the URN from the raw string:

// packages/tool-server/src/tools/debugger/debugger-service-ref.ts:89
return `JsRuntimeDebugger:${params.port}:${deviceId}`;

Every other place that looks a device id up folds case first:

• device-alias.ts:86,92,97 — logicalKeyedConnectIds add/has/delete all .toLowerCase()
• device-services.ts:213,215 — deviceIdOwningUrn lowercases both sides, with a written rationale
• reaped-sessions.ts:87 — key() lowercases, pinned by reaped-sessions.test.ts:40 ("iOS UDIDs come back uppercase from simctl and lowercase from some tool args, and every id lookup in the stop tools already compares case-insensitively")

So 6DBF83B4-… and 6dbf83b4-… are one device to the alias map, the teardown scoper and the breadcrumb store, and two devices to the registry. selectTarget compares with === (target-selection.ts:54) but falls through to the single-device fallback, so both spellings resolve to the same target and each mints its own JsRuntimeDebugger node — a second CDP socket to Metro, a second bound loopback console-log server, and a second log file, all for one device. On Metro that also evicts the first session, since its inspector proxy allows one debugger per device (#909).

The two conventions meet in the breadcrumb store: one folded slot serves two unfolded sessions, so a record filed by the second spelling's session is what the first spelling's reader gets. Reproducible directly against the store — record 6DBF83B4-… with keptAt=P1, then 6dbf83b4-… with keptAt=P2: the supersede path unlinks P1 and a read under the upper spelling returns a note naming P2.

Folding in the URN builder looks like the consistent fix — it is the one holdout, and a device id is not case-sensitive on any platform argent supports (simctl UDIDs, adb serials, chromium-cdp-<port>).

Not caused by #705: debugger-service-ref.ts:89 is unchanged by it. Found while reviewing it — that PR's breadcrumb store is where the asymmetry becomes observable, but the duplicate sessions predate it.