Release v0.6.36: offload notification registration and mitigate Viber paste timing

Release/README.md

@@ -1,18 +1,17 @@
 # Latest Windows release
 
-## v0.6.35 — 2026-11-04
+## v0.6.36 — 2026-11-04
 
 ## Changes
-- Reduced resident memory: the input predictor and correction worker share immutable dictionaries. Compact indexed blocks retain all 140,939 bundled words and their frequency ranks while reducing the EN/UK/RU lookup payload from approximately 3.06 MiB per copy to 0.90 MiB shared. Custom UTF-8 dictionary files, duplicate handling, frequency limits and personal dictionaries remain compatible.
-- Closing clipboard history releases its duplicate row list and Windows list strings, retains selection/scroll identifiers for reopening, and rejects late search results and timers. Preview caches now have payload budgets of 1 MiB for thumbnails and 256 KiB for formatted text, plus the existing item-count limit. Visible entries are retained even if an unusually large viewport exceeds a budget.
-- PNG decoding reads the existing input buffer without making another full PNG copy. DIB decoding allocates its temporary BMP buffer once at the required size.
-- Idle maintenance releases diagnostic context older than 30 seconds and expired Undo text after its existing two-minute eligibility window, without interrupting an active operation. Completed sync buffers are released on success, failure and stale responses; unresolved conflicts remain available.
-- No shortcut, default, on-disk history format or screenshot-editing behavior changed. The v0.6.34 four-direction movement cursors remain included.
+- Notification registration now runs in the existing KSwitcher.Screenshot.exe helper. Startup launches it asynchronously and checks its exit status without blocking keyboard work; update notification registration waits in its background worker. Keep all three executables together. UI Automation and notification activation routes remain unchanged.
+- Viber paste now allows 75 ms for clipboard notifications to settle before sending Ctrl+V, then rechecks input, focus and clipboard ownership. This mitigates a suspected clipboard timing race: an intermittent corrupted Cyrillic paste was reproduced on the prior build. It is not a guarantee that every Viber timing issue is resolved; no blind paste retries or relaxed selection checks were added.
+- Added regression coverage for an explicitly selected middle range followed by a repeated shortcut without selection, preserving surrounding text, trailing spaces and the original clipboard. Updated the integration fixture to initialize shared predictor dictionaries.
+- Added reproducible standalone COM, UI Automation and Shell memory measurement tools. No shortcuts, settings or defaults changed.
 
 ## Validation
-The full `build.ps1 -Stage -Test` suite covers complete bundled dictionary word/rank equivalence, malformed/custom dictionary boundaries, 105 sync checks, shared wheel batching, 304 native regressions, 559 input/UI checks, screenshot editor UI and PNG saving, isolated screenshot registration, 48 updater regressions and updater progress UI checks. Cache assertions now run once outside the exception helper rather than being repeated inside it. All three executable versions and all 12 package manifest file sizes/SHA-256 hashes were verified; no personal data is packaged.
+Full staged build and tests cover dictionary equivalence for all 140,939 words, 105 sync checks, wheel batching, 304 native checks, 559 input checks, screenshot UI/PNG/registration, 48 updater checks and updater progress UI. Conversion integration passed 43 assertions. Eight actual Viber conversions passed on the development build, including two complete long-text layout cycles and selected/repeated middle-word conversion; test messages were not sent.
 
-Controlled local measurements of the final dictionary step, after equal restart/wait periods with three dictionaries enabled, reduced private resident memory from 7.08 to 4.65 MiB and private committed memory from 8.34 to 5.70 MiB. These compare consecutive development builds, not the published v0.6.34 with the full cumulative change. Randomized lookup tests were approximately 20% faster. Results depend on settings, workload, Windows and allocator state; these are not fixed memory limits. The 0.90 MiB dictionary figure excludes process/allocator overhead.
+A controlled fresh-launch before/after comparison (five private-working-set samples each, same installation and settings) measured 4.69 → 4.18 MiB, about 0.51 MiB saved. This is one local comparison, not a fixed memory limit or a long-term idle guarantee. The prior compact dictionaries and cache/idle memory reductions remain included.
 
 ## Upgrade
 Choose **Update** in the tray menu. Legacy v0.3.0–v0.4.1 clients first receive the v0.4.2 compatibility bridge, then the current full release. v0.2.3 and earlier require manual installation.
@@ -20,4 +19,4 @@ For manual installation, exit KSwitcher, close screenshot editors and extract th
 
 Download **KSwitcher-win-x64.zip** with **SHA256SUMS.txt**. Keep all three executables together. The package contains no personal settings, credentials or clipboard history.
 
-Publish stable `release-vX.Y.Z` tags matching the executable version. Keep `v0.4.2` as the last legacy tag.
+Publish stable `release-vX.Y.Z` tags matching the executable version. Keep `v0.4.2` as the last legacy tag.

Release/SHA256SUMS.txt

@@ -1 +1 @@
-6e36a582eec59b3cc4f0f9d6b3e05cd699371c05cae30a95c6eae51a838f7bfe  KSwitcher-win-x64.zip
+794d74113bddf13ffee7317cadee9c472bca3b84b3361796db58ff8023f66571  KSwitcher-win-x64.zip