Optional Recursive Mutex in Asio Integration - #1846

…s and the corrected Visual Assist reading on BPR-0738

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Six candidates at app-release-e4fdb0ea-2026-09-11.4 on the tests-disabled
signed bundle built at 3543824 (BPR-0646 and BPR-0720 re-asked on
NVIDIA/stdexec#1846; BPR-0734, BPR-0737, BPR-0738, BPR-0739), ready and
protected at sequence 341; no host started. The six open cards carry the
v008 pointer and BPR-0593 the d18 frontier.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

Every verdict is recorded against its own candidate and fingerprint; all
six match the v008 identity file, so no verdict is stale.

BPR-0646: the handle is unjudged for the third round, and not because of
the pull request. An epoch is this install's own observation and nothing
enumerates a pull request's earlier updates, so a first open yields one
epoch for any pull request; the local catalogue holds exactly one row for
NVIDIA/stdexec#1846, minted five minutes before the verdict. BPR-0743
holds the owner's choice between importing history and rewording the
criterion.

BPR-0720 and BPR-0738 are one requirement in two notes: a directive line
paints grey end to end, where the reference greys the directive, paints
its ALL_CAPS argument as a macro and its include path as a string.
BPR-0744. The include-path question is answered; the alias italic and
the CE8AD4-for-BD63C5 substitution are not, and are recorded as still
open rather than read into "most things are good now".

BPR-0734: the one-pane map is the link map at its limit, not a new
element, and the RFC's non-goal is relaxed for that direction. BPR-0745
redraws the state. Block order and where the choice lives stay
unanswered and are asked again.

BPR-0739: the seam is accepted, the cursor is not. The highlight is
painted only from underMouse(), so the handle held Qt's pointer state
while the cursor stayed an arrow, which refutes the d17 cause. BPR-0746
carries the cursor-delivery hypotheses and the check that discriminates
them on the real application. AC1's cocoa-probe sentence is corrected:
that log has the pointer 1164 px from the handle and prints FAILED.

All five cards stay parked on human-review.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The signed-out obstacle does not exist. Run rather than reasoned about: GET
/repos/NVIDIA/stdexec/issues/1846/timeline is HTTP 200 with no credential,
against the 60-per-hour unauthenticated budget. The timeline names both force
pushes with GitHub-side ids and the head each moved to, and one of those heads
-- 6ffbd0ac -- appears nowhere in the commits list, so this card's claim that
commits cannot reconstruct the updates that matter is now demonstrated rather
than asserted. That head is still fetchable, so an import would not list
updates it cannot then diff.

Three limits came out of the same response and the design is built around
them: the REST force-push event has no before-commit (GraphQL has it and is
refused anonymously), array position is not an ordering key and `committed`
dates are author dates, and an ordinary push is not an event at all.

Decision 1, the derivation: a digest over the authority scope and the update's
head OID, carried beside review_epoch_id. Not the event node_id, which exists
only for force pushes and which catalog_database.cpp:86-88 excludes by design;
and not the base/head pair this card offered as the fallback, because the base
is time-dependent even for a live read.

Decision 2, the ordinal: not redefined. Redefining it cannot buy portability
while Incarnation stays in the id, two of its seven serializations are already
outside the catalogue's reach (git refs on disk, MCP bindings and agent launch
JSON already handed out), and density is an adapter guarantee that history
paging and alias resolution have come to depend on.

Decision 3 is the honest limit: identity can be made stable, population cannot
be made equal. A head observed live and then force-pushed over is recoverable
by nobody, which is precisely why the identity must not be the ordinal.

Decision 4 is where it exceeds the card. A base OID is mandatory and
origin-blind, and the timeline supplies none for a historical update.

So the conditional fires, on kind rather than size: identity, base OID,
observed time and the schema version all move, four of the six things
RFC-BPR-0003 pins. Nothing is implemented. The ask is one question -- its own
RFC, or an amendment against the closed one -- and it is on the card.

Three of the card's anchors were wrong and are corrected in place: the
anonymous admission header's path, the ordinal id write line, and the
sixteen-hex-digit ref encoding, which is in installed_git_process.cpp and not
installed_git_cache.cpp.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>