fix: normalize peer address before duplicate-identity comparison - #48

On two fixed-MAC peers both running central + peripheral, the same
physical peer is represented by two address strings: the peripheral
(GATT) path stores "dev:AA:BB:.." (BlueZ D-Bus device-path form) while
the central (scan) path carries the bare "AA:BB:..". The
_check_duplicate_identity comparison (and the v2.2 scan-loop comparison)
saw "dev:AA:BB" != "AA:BB", concluded a false Android MAC rotation,
rejected the connection, and detached the peer interface after the grace
period. The data path then dropped: discovery announces routed to 0
peers and the peer never appeared in the destination table.

Strip the "dev:" prefix and case-fold before comparing, so the same
physical MAC compares equal regardless of which form it arrived in,
while genuinely different MACs (true MAC rotation) still differ.

TDD: test_ble_dup_identity_mac_normalize.py reproduces the false
positive (RED before, GREEN after) and guards that true MAC rotation is
still rejected. Bind the new _normalize_address helper in the
zombie/blacklist test harnesses that call the real method.

The normalized same-address branch in _select_peers_to_connect (dev:-prefix
peripheral form vs bare central form for the same fixed-MAC peer) is now
pinned by TestScanLoopSameAddressRegression: the skip decision (interface
already exists, so the peer is not re-added as a 'MAC rotation') is asserted,
with a stale-interface (not-in-self.peers) state, and a contrast test proving
that the skip is caused by the branch rather than another gate.

Also documents that the same-MAC reconnect path is the standard top-level gate
(address in self.peers), not this branch.

Android (AndroidBLEInterface) and iOS subclass the shared BLEInterface and
wire on_duplicate_identity_detected -> _check_duplicate_identity with no
override. Their drivers produce bare, same-case MACs (never the BlueZ dev:
prefix that is the headline Linux fix), so the normalizer must be a no-op for
the common mobile case and must not disable true MAC-rotation rejection.

Add TestBareMacMobileSafety to pin that contract on the real method:
  * bare same-case MAC, alive  -> not a duplicate (unchanged)
  * bare different MAC, alive  -> still rejected (guard holds)
so a future edit cannot silently alter Android/iOS behavior.

…nitpick)

The TestScanLoopSameAddressRegression class lives in tests/test_v2_2_mac_sorting.py
(not test_ble_dup_identity_mac_normalize.py); fix the pointer in the code comment.

The v2.2 same-identity normalization in _select_peers_to_connect is pinned
by tests/test_v2_2_mac_sorting.py, but that file is excluded from the
integration coverage run (--ignore=...), so codecov/patch sees the changed
scan-loop lines as uncovered and fails.

Add tests/test_scan_loop_normalize_coverage.py that drives the REAL
_select_peers_to_connect from an included module, pinning each branch
(rotation-alive skip, rotation-dead reselect + cleanup, same-normalized
address skip) plus the _normalize_address falsy-input branch. Verified:
the three changed BLEInterface regions (check_duplicate_identity,
scan-loop norm, _normalize_address) now show 0 uncovered lines under the
CI-equivalent coverage run.