v11.23.0

Includes everything in v11.22.12 plus the full Phase BM landing
plus six infra root-causes diagnosed + fixed during BM.5 verification
(each one a real latent bug, none xfail'd).

Phase BM (dissolve the dual-SQL date-filter form)

Replaces the pre-BM dual SQL form (analysis-level TimeRangeFilter
on QS + {date_filter} template + app2_date_filter() for App2)
with a unified <<$pXxxDateStart>> / <<$pXxxDateEnd>> dataset
parameter pushdown across both renderers. Same <<$pX>> pattern
Phase Y used for categorical / slider filters. Both renderers narrow
at the DB; the QS day-edge quirk + visual_identifier= test dance
artifacts dissolve by construction.

Coverage:

• L1 dashboard (8 datasets) — shared pL1DateStart /
  pL1DateEnd across drift, ledger_drift, overdraft, limit_breach,
  todays_exceptions, transactions, drift_timeline,
  ledger_drift_timeline.
• Executives (3 datasets) — pExecDateStart / pExecDateEnd
  for transaction_summary, transaction_daily,
  account_summary_active.
• L2 Flow Tracing (4 datasets, BM.5 sweep) — per-sheet pair
  (pL2ftDate* for Rails-on-postings, pL2ftChainsDate* for
  chain_instances, pL2ftTtDate* shared across tt_instances +
  tt_legs). Pre-BM L2FT's App2 pickers were UX lies (widget shown,
  no SQL narrowing); now both renderers narrow.

Dead machinery removed:

• build_dataset.app2_date_column kwarg + {date_filter} slot.
• common/sql/app2_filters.py::app2_date_filter() (file rewritten
  to expose only universal_date_range_clause).
• Analysis.default_universal_date_range field + the BL.2 bind-
  layer prepop helper.
• DateView.emit_app2_date_from / emit_app2_date_to.
• App2 universal-range hidden date_from / date_to block — each
  picker now renders as its own ParameterDateSpec with a
  param_<name> URL key, symmetric with other categorical
  pickups.
• Both bg5 [qs] xfails (test_bg5_transaction_volume_kpis_* +
  test_bg5_money_moved_kpis_*) — the day-edge quirk dissolves by
  construction so the strict=False xfails are gone.

BM.5 infra root-cause fixes (each diagnosed, not xfail'd)

1. xdist dist=loadgroup never fired under bare pytest -n N:
   the conftest's pre-BM check matched only dist == "load" but
   xdist's default-to-load happens after our hook runs. Bump now
   matches "load" OR "no". Fixes
   test_invariant_three_way_agreement[anomaly] DROP-race between
   two xdist workers racing on the same iagree-prefixed matview.

2. Freshness oracle deadlock on filter-control-free sheets:
   emit_visual_data_fragment skipped data-bound-params when no
   param_* / filter_* URL keys were present (L2 Exceptions
   sheet shape). bootstrap.js's beforeRequest always stamped
   data-requested-params="{}", leaving the oracle deadlocked on
   req="{}" === ren=undefined. Always emit ("{}" when empty).
   Fixes test_bg6_l2ft_exceptions_*[app2].

3. Oracle DPY-4008 from :MI/:SS inside TO_DATE format
   literals: oracledb's pre-execution bind-name scanner doesn't
   respect SQL string-literal quoting and tripped on
   'YYYY-MM-DD"T"HH24:MI:SS'. Sidestep with
   TO_DATE(SUBSTR(<param>, 1, 10), 'YYYY-MM-DD'). Documented in
   docs/reference/oracle-19c-constraints.md. Fixes the entire
   Oracle leg of the BM date pushdown.

4. Spine Invariant.detect() portable cursor helper: pre-BM
   the 12 detect methods called conn.execute(sql).fetchall()
   relying on sqlite3.Connection's convenience shim. oracledb +
   psycopg only support DB-API 2.0 cursor-then-execute. New
   common/spine/_db.py::fetch_all(conn, sql) swept across all
   12 modules. Fixes 4 test_spine_live_agreement tests on the
   Oracle/AWS variant.

5. Bun lightningcss DLOPEN race on parallel cell pytest
   invocations: the tailwind drift check fires Bun which
   extracts a native lib to /tmp; 13 parallel cell pytest sessions
   racing on the same extraction crashed cells with
   ERR_DLOPEN_FAILED. Runner now sets RECON_GEN_SKIP_TAILWIND=1
   on per-cell layer subprocesses (the unit prelude already runs
   the gate cleanly once at session start). Fixes sp_sl_lo
   app2-layer crash.

6. Docker PG max_connections bump: postgres:17-alpine
   defaults to 100. Per-cell pytest -n auto × ~16 xdist workers
   × per-fixture App2 pool size 10 = ~160 connections in-flight.
   Bump to 300 via .with_command(). Fixes sp_pg_lo PoolTimeout
   cascade.

Test surface

• 3,582 non-e2e tests green throughout.
• sp_pg_aw browser layer green end-to-end (the major PG/AWS
  variant) — proves the BM core + 5 root-cause fixes hold under
  live AWS load.
• Remaining sasquatch+AWS variant failures (sq_pg_aw / sq_or_aw /
  sp_or_aw) filed as Phase BN (sasquatch seed-vs-BM-date-window
  • Oracle/AWS load-scaling). Independent of BM-the-mechanism.