Validate audit windows and isolate dashboard count reads

docs/ui-consistency/ScadaBridge-verification.md

@@ -63,3 +63,29 @@ Parent review found the raw-JSON escape-hatch error was not described by its com
 Commit915a8a59 was created before the author checked the focused result carefully. Its raw-json-association-green.log actually contains98 passes/1 failure and remains unchanged as historical evidence. The fixture incorrectly expected valid JSON object syntax to erase the existing unresolved-schema error. This was a test-premise error; production validation must retain that error.
 
 The corrected regression now asserts invalid syntax is linked, valid object syntax still retains the unknown-declared-type error, and then valid JSON null clears invalidity under InboundApiSchema's existing explicit present-null-is-allowed contract. All prior relationship/domain-name assertions remain. No production source or validator changed in this correction. Focused ParameterValueForm/Monaco/TreeView result:99/99 passed, zero skips/failures (raw-json-association-corrected-green.log). Parent accepted focused verification after the prior1559 full baseline. This paragraph supersedes the premature green claim above; it does not rewrite the red evidence.
+
+## SB04–06 review batch
+
+SB04's UTC-labelled native inputs now belong to AuditFilterBar. The reusable DateTimeRangeFilter remains unchanged: its DateTime? callback intentionally cannot distinguish invalid raw text from an empty bound. The parent approved using its same native two-field geometry/labels locally so the audit caller can keep correction strings. Valid literals are tagged UTC without browser/server-zone conversion, including DST-date literals. Intentional nullable bounds remain supported; invalid/inverted input returns before OnFilterChanged, preserves strings, and links a recovery error. The query model, presets, filters and backend remain unchanged. New error uses shared field-error s-bad (#b42121), not a local color override. Clear resets raw strings/error together with existing model reset.
+
+SB05 dashboard facts settle independently after each read with individual loading/known-zero/unavailable state. Repository reads remain sequential because some adapters share one scoped DbContext. A failed read does not block later reads or replace available facts. Each named region's Retry calls only its read. No domain service, aggregation, polling or authorization change. SB08/SB10/SB11 still own dashboard hierarchy, destination policy and operational-summary changes.
+
+SB06 adds contextual Dismiss [title] names to toast close controls and preserves timing/live-region/disposal behavior. Two-toast test proves one dismissal preserves the other message.
+
+Evidence sb04-06: red-utc-toast.log6 intended failures; red-dashboard.log6 intended failures. Full Release suite1575/1575 passed, zero skips/failures (full-suite.log/sb04-06.trx). Final audit tests20/20 pass under UTC, America/Indianapolis and Asia/Tokyo; custom, DST, invalid/inverted, correction, nullable and existing preset/filter cases remain green. Only a shared semantic error class and stronger false-zero assertion followed the full run; final-focused.log covers the final audit/dashboard/toast state.
+
+First8-case browser matrix retained in browser/results.json. UTC literals/inverted guards passed all4 cases (New_York/Tokyo browser zones ×1440/390). Immediate dashboard retry and one toast count were read before InteractiveServer completion and reported false. The retained confirmation adds outcome-based waits; no source state logic changed for these measurement failures. An intermediate harness quoting typo is retained as confirmation-harness-error.log. Final bounded confirmation8/8 passes (browser-confirmation/results.json): same UTC boundaries, input text retained on inverted range, partial-dashboard Unavailable→known-zero retry with successful facts retained, and named target-only dismissal. All captured dashboard/toast document widths equal viewport width; visual inspection confirms existing density and hierarchy are retained for later SB08 changes.
+
+Fixture changes under umbrella validation/scadabridge-ui/host supply only empty read projections and one deliberate first API-key-read failure. All non-read operations reject; no Save/test-send/verify/trust/deploy was performed. Direct actual component composition does not establish application authentication or live backend behavior. Original G1 evidence is unchanged.
+
+This batch's production scope is4 Razor/C# source files (AuditFilterBar pair, Dashboard, ToastNotification) plus3 test files; frozen hashes are sb04-06/source-freeze.json. No SB07+ source changes or batch commit before review.
+
+### SB04–06 concrete peer corrections
+
+The original bUnit invalid-string test did not prove native partial-date behavior. Actual Chromium keyboard entry reproduced value empty + badInput true and the prior onclick Apply emitted an unbounded filter (native-partial-before.json). AuditFilterBar now owns one native form; its only production callsite in AuditLogPage has no surrounding form. Apply submits, Clear is explicitly type=button, and native constraint validation blocks incomplete values before the Blazor submit handler. Tests submit the form instead of directly invoking a button click; all existing filter assertions remain. The input uses step=any to avoid unintentionally rejecting valid seconds. Intentionally cleared optional bounds remain accepted.
+
+The native after proof uses actual segmented keyboard entry, not programmatic invalid value assignment: partial input is blocked, focus returns to From, then a genuinely cleared optional bound submits null. Explicit aria-invalid=false initially masked native badInput in Chromium AX. The caller now emits aria-invalid=true only for its own range errors and otherwise omits it; native AX reports invalid=true for a partial native date. Before/after AX proof is retained with native-partial-after.json and native-partial-aria-before.json. No JS validity workaround.
+
+Dashboard initial reads alone were sequential, but an early Retry could overlap the next repository read. All initial/retry operations now share a component-owned async gate. The renderer claims each fact's pending state before its first await, coalescing duplicate target requests. A deterministic delayed-Connections/early-Sites-retry regression reproduces the prior3 Sites invocations before Connections completed (expected1); final code waits and makes exactly one recovery read. Counts remain independent, and successful facts are not reloaded. The first async test attempt had CS4014 in NSubstitute verification; compile output is preserved separately, then meaningful red/green evidence follows.
+
+Reviewed full Release suite1576/1576 passed, zero skips/failures (reviewed-full-suite.log/sb04-06-reviewed.trx). The only following production change was omitting the false ARIA override; final focused31/31 and final20 audit cases in each of UTC/Indianapolis/Tokyo pass. Native Chromium focused after proof passes including AX invalid=true. These correction checks are narrow technical regressions, not another broad screenshot/polish round. No batch commit or SB07+ expansion until parent accepts the corrected frozen source.