USE_BLUEDROID_GATT_SERVICE_TABLE compiles the Bluedroid materializer and
nothing else; the old name read like a cross-backend consumer capability
(rp2 always materializes and ignores it).
Each backend arm now has a codegen-emitted USE_BLE_GATT_BACKEND_<NAME>
define, making the alias ladder order-independent and letting the proxy
capability constants describe the backend instead of the platform (a second
esp32 backend carries its own facts). The legacy raw-gattc engine is
selected by USE_BLE_CLIENT_LEGACY_ENGINE instead of platform identity, so
esp32's future move to the neutral engine is one line per language. The
address-scoped maintenance calls key on the stack define they wrap.
Connect-timing ownership (the backend integrates with its platform's
scan/connect arbitration), the asynchronous-completion rule for accepted
ops, and the reserved error ranges with the wire-frozen values - all learned
the hard way by the first out-of-tree backend.
has_legacy_nodes_() lives next to the invariant it derives from; both
bailout outcomes reach one promotion loop; the carve assert relaxes to the
descending chain the comment claims; the notify op string names the IDF
call like its siblings.
The static_assert enforces what the carve comment claims; the node header
now states that aborting inside on_connected() can leave on_disconnect
unpaired; the passive-proxy test comment names the current gate; the
pending-regs comment lands on the right member (lost to an aborted batch).
A failed search still poisons the whole discovery, but an allocation or walk
failure is table-specific: on a mixed client the legacy nodes keep the link
and only the gatt nodes skip that connection.
The write action's own tag left Automation::TAG with zero readers;
automation.cpp existed only to define it. Also notes the unpaired legacy
disconnect trigger on the failed-discovery path and re-anchors the
gatt_nodes_ comment.
One ConnectBackoff shared by both engines replaces the forked hold-off
fields and constants (the duration is derived from the failure count). The
pending-registration scan lives once, sized by its own bound instead of the
node counter. The failure-suppressing SEARCH_CMPL arm moves to the call site
so the dispatcher stays void, notify-state logging is one helper on the
neutral wording, and the op-check helper takes the backend's spelling.
Failed notify enables, reads, writes and unsubscribes now leave the same
breadcrumbs as the neutral engine; the shared-completion guarantee moves to
the node contract header.
set_enabled(true) resets the hold-off (neutral-engine parity), and every
failing build() return now resets the counts so a failed build can never
present a non-empty view.
A discovery the bridge classifies as failed now suppresses the legacy
SEARCH_CMPL fan-out, so user automations never observe the doomed link as a
connection (neutral-engine parity, including the mid-fan-out node abort).
Also documents that same-handle subscriptions from mixed node kinds are
unsupported during the migration window.
A peer that reliably fails discovery (walk error, allocation failure) would
otherwise reconnect on its next sighting forever, holding a controller slot
each cycle - the neutral engine already backs off on this path.
Neutral fan-out runs before the legacy nodes so on_connect triggers observe
resolved gatt nodes (the neutral engine's order). Pending notify
registrations move to a fixed array (no post-setup heap). Gatt nodes also
join nodes_, collapsing the twin state loops. One define now means 'the
neutral node surface is compiled in' on both engines. The service count
moves into BluedroidServiceTable, free() resets it, bridge ops log failures,
the write action carries its own tag, and the automation headers drop
includes that left with the write actions.
Platforms whose BLE stack owns a connection budget (esp32, rp2) report an
overcommit exactly once through that stack; the neutral cap check covers
future backend platforms without one. The write action logs unmatched
completions while a chain is parked, and the schema-dump gap is documented
as deliberate.
One implementation replaces the raw-gattc one and runs on esp32 and rp2
alike; user YAML is unchanged. The node resolves the V1/V2 variant from the
service table, owns its CCCD write, and keeps the drop-the-link-between-polls
protocol.
The raw-gattc and neutral twins collapse into ble_write_action.h, registered
through register_gatt_node on both engines. On esp32 this is the in-tree
exerciser of the node bridge; builds without a ble_write action compile no
bridge code.
BLEClientFeatures (gatt_node / raw_gattc / security) with a per-platform
provider set, capability-named validation errors, NODE_BLE_CLIENT_SCHEMA +
register_gatt_node for migrated components, and the security actions gated on
the feature instead of the platform. esp32 auto-loads bluetooth_connection
for the shared materializer (compiles empty without a neutral node).
Nodes registered through register_gatt_node() get the neutral surface:
service table at SEARCH_CMPL (honoring the search status the base ignores),
translated completion/notify fan-outs, pairing results, and teardown via
on_disconnect_complete. Bridge-initiated notify registrations are intercepted
before the base's automatic CCCD write - the CCCD is the node's job on the
neutral contract. Compiled only under USE_BLE_CLIENT_GATT_NODES; builds
without a neutral node are unchanged.
BluedroidServiceTable carries the two-pass build verbatim in its own TU; the
backend embeds one instance. ble_client's esp32 engine becomes the second
consumer next.
The neutral callbacks (plus a new on_pairing_result hook) live unconditionally
in ble_client_node.h; the raw esp32 surface stays beneath them for unmigrated
nodes. Adds find_descriptor beside find_cccd and pair()/unpair() on the
neutral client. Drops the node 'client' pointer and address, which nothing
reads.
J. Nick Koston