fix: make public claims precisely true, plus three small fixes

[LucaGerlich]

First PR from the post-launch audit (2026-10-11). Corrects every public claim the code contradicts and fixes three small bugs. No feature changes.

Claims (the product's pitch is precision about Input Monitoring)

• Structured data on the site still said "never key codes" — the exact sentence corrected in August. The app reads the HID usage code to detect modifiers, then discards it.
• "Nothing leaves your Mac" on the Welcome screen ignored Sparkle's automatic daily update check. Welcome, README, site and SECURITY.md now describe the update check and update downloads.
• "⌘-Tab never changes your layout" was false: Tab is not a modifier. Now: a lone modifier is ignored; any other key counts as typing.
• "Right the moment you start typing" → it switches about 0.4 s after typing starts, so the first character or two can land in the old layout.
• SECURITY.md: the usage code is a local constant; the storage list is complete (only mapped keyboards are stored, plus Sparkle's bookkeeping); redaction runs when an entry is recorded; Sparkle checks a valid code signature, not notarization.
• CHANGELOG: only update archives are signed, not the appcast itself (that changes in a later PR).
• The comparison FAQ now names Keychange, the closest alternative.
• The historical pre-release audit gets a correction note and no longer names a work account's team ID.

Fixes

• Debug log flood: while the HID monitor could not start, the same error was appended every second until the 200-entry log held nothing else. Each distinct failure is now logged once; a different failure is still logged.
• F6 treated as a modifier: usage 0x3F is F6 on the HID keyboard page (it is Fn only as a Carbon virtual keycode), so a lone F6 never triggered a switch.
• An always-true test assertion (String != nil) now checks the actual warning text.
• README gains the Homebrew install section that was stranded on docs/public-polish.

Test plan

• RED first: the F6 test and the log-flood test failed against the old code for the expected reasons (usage 63 → modifier; 5 identical error entries).
• 38 unit tests + 2 UI tests pass locally.
• No compiler warnings in app or test sources.
• JSON-LD parses; all 7 FAQ questions and answers are identical between JSON-LD and the visible page.
• Independent review of every changed claim against the code.
• CI green.
• After merge: check the live site (GitHub Pages, ~1 min).