fix(cli): name --udid when a device identity is passed to --device

[NicolasBataille]

Closes #2064

Summary

--udid is the only selector that pins one simulator when several share a name, and it was
effectively undiscoverable: help commands listed no device selector at all, and passing a UDID to
--device answered with a literal lookup failure that neither named --udid nor hinted that a
UDID-taking flag exists.

Before:

$ agent-device snapshot --platform ios --device 204BFFD9-9644-4830-B2C1-1B946597A07C
Error (DEVICE_NOT_FOUND): No device named 204BFFD9-9644-4830-B2C1-1B946597A07C
Hint: Verify the target device is booted/connected and selectors match.

After:

Error (DEVICE_NOT_FOUND): No device named 204BFFD9-9644-4830-B2C1-1B946597A07C
Hint: 204BFFD9-9644-4830-B2C1-1B946597A07C is the id of "iPhone 17", not its name.
      Did you mean --udid 204BFFD9-9644-4830-B2C1-1B946597A07C?

Two changes:

1. resolveDeviceByName answers the flag mistake. When the --device value is the id of a
   listed candidate it names that device and the flag the value belongs to — --udid for Apple,
   --serial for serial-addressable platforms. When it merely has UDID shape (8-4-4-4-12 hex) but
   matches no listed device, it names --udid anyway, because a shut-down simulator is still a
   wrong-flag answer. An ordinary unknown name keeps the generic hint untouched.

   This mirrors, rather than invents, the answer assertSelectorFlagMatchesPlatform already gives
   a few lines above for a mismatched identity flag (--udid on --platform android).

2. help commands renders a Device Selection section for the selectors every command
   accepts: --platform, --device, --udid, --serial, --session.

   Device Selection (accepted by every device command):
     --platform apple|android|...  Platform to target (`apple` aliases the Apple automation backend)
     --device <name>               Device name to target (a UDID belongs in --udid, a serial in --serial)
     --udid <udid>                 Apple device or simulator UDID; the only selector that pins one
                                   device when several share a --device name
     --serial <serial>             Android device or Vega VVD serial
     --session <name>              Named session
   

   Deliberately a new help section rather than a GLOBAL_FLAG_KEYS addition: those keys mark a
   flag as supported by * in buildOptionSpecs and move it from supportedFlags to globalFlags
   in explain. These selectors reach device resolution, not the CLI envelope; they stay in
   COMMON_COMMAND_SUPPORTED_FLAG_KEYS, so the option schema, explain output and MCP surface are
   byte-identical.

--udid and --device also get sharper one-line descriptions, which is where an agent reading the
new section actually learns the distinction.

Tests

• packages/kernel/src/device-selector-flags.test.ts — the existing home for selector-flag
  mistakes — gains four cases: a UDID and a serial passed to --device (each naming the right
  flag and the device it identifies), a UDID-shaped value no listed device carries, and an ordinary
  unknown name keeping the generic hint. All four were observed red against the pre-fix code.
• src/cli-schema/cli-help-topics.test.ts gains a Device Selection section assertion. The
  existing "only global flags in its global flags section" test slices from Global Flags: to
  Configuration:; the new section sits above Global Flags:, so that slice is unaffected.

pnpm check:quick, pnpm test:unit (8092 passed), pnpm check:command-docs and
pnpm check:agent-guidance are green.

Live check

Run from the built clone against a local simulator set with two simulators named "iPhone 17":

• --device <udid> → the new identity hint naming --udid and "iPhone 17".
• --device 11111111-2222-3333-4444-555555555555 (no such device) → the shape hint naming --udid.
• --device "iPhone 99" → unchanged generic hint.
• help commands renders the new section.

What a maintainer might push back on

• The UDID regex. It is deliberately the fallback: the primary signal is exact id-equality
  against the candidate set, which needs no shape knowledge and gets --serial right for Android.
  The regex only covers a device the current inventory does not list. Dropping it would still fix
  the reported case; keeping it covers the shut-down-simulator variant the issue's workaround
  section implies.
• --device accepting a UDID outright was the issue's other suggested fix. Not taken: it would
  make --device a two-meaning flag and quietly duplicate --udid/--serial's platform routing,
  including the assertSelectorFlagMatchesPlatform cross-check. Naming the right flag is smaller
  and keeps one meaning per selector.
• Section placement and title. Device Selection (accepted by every device command) is long;
  a shorter title works if it does not read as "global", which is exactly the distinction that
  keeps the option schema unchanged.
• Selector set. --target is also a common selector and was left out to match the issue's list.
  Adding it is a one-line change to DEVICE_SELECTION_FLAG_KEYS.

[thymikee]

[P1] Remove the partial UDID-shape heuristic and keep the hint grounded in observed device identity. The exact candidates.find(device.id === deviceName) path already fixes #2064 for every discovered Apple/serial device and reuses the kernel’s platform facts. DEVICE_UDID_SHAPE then reconstructs Apple identity syntax with a UUID-only regex: it excludes the modern physical-device UDIDs already used in this repo (00008150-001849640CF8401C), can misclassify an arbitrary UUID-shaped device name, and needs a long comment to justify a shutdown-device case that current inventory may still list. Either delete this fallback (the smaller fix the PR body already identifies) or move complete Apple identity recognition behind an Apple-owned typed interface with tests for simulator and physical formats. Keep the exact-candidate --udid/--serial hint and the help section. Exact-head CI is also still absent, so do not apply ready-for-human yet.