feat: give the touch stack three distinct verdict screens

[buzzromain]

Why

NBGL's check-result screen shared one title, "Valid Secret\nRecovery Phrase", between two different outcomes: a phrase that is well formed but does not match this device's seed, and a phrase that matches. Only the paragraph underneath said which. The two-button screens already made this distinction: ux_bip39_nomatch_flow and ux_bip39_match_flow (src/bagl/ux_nano.c) carry two different titles, "doesn't match" and "is correct", on two separate flows. NBGL's invalid-phrase screen also gave no advice, where the two-button invalid flows share a line telling the user to check length, order and spelling.

This is the single most security-sensitive screen in the application — the one screen where a fault injection has been shown to be able to flip the displayed result. An ambiguous verdict on it is the most expensive kind of interface bug this project has.

What changed

Two additions to src/common/ui_strings.h: UI_STR_NBGL_RESULT_NOMATCH_TITLE and UI_STR_NBGL_RESULT_INVALID_ADVICE. display_check_result_page() (src/nbgl/ui.c) now indexes a three-title array by the same 0/1/2 outcome it already computed for the verdict icon (previously icons[result + seed_match]; that expression is unchanged, just factored into a named outcome variable also used for the title), instead of pulling a shared title from the two-row body-text table. The invalid case's advice line goes into nbgl_pageInfoDescription_t.centeredInfo.text3.

No two-button file changed — those flows already worked this way.

An SDK pitfall found before it shipped, not after

text3 only renders as its own line under the LARGE_CASE_GRAY_INFO style. Under LARGE_CASE_INFO — what this screen used before — nbgl_layoutAddCenteredInfo() (SDK) maps both text2 and text3 onto the same underlying centeredInfo.description field, so setting text3 would have silently replaced the verdict's own body text rather than adding a line under it. Confirmed by reading the SDK source, then reproduced under Speculos on stax before applying the fix: the body text disappeared, replaced by the advice line, the moment text3 was set under the old style. The style is LARGE_CASE_GRAY_INFO in this PR.

Wording

The new nomatch title is "Mismatched Secret\nRecovery Phrase" — chosen to keep the exact "<Adjective> Secret\nRecovery Phrase" template the other two outcomes already use ("Invalid Secret", "Valid Secret"), including the same line break, after "Secret" rather than after "Recovery". An earlier draft, "Recovery Phrase\nDoesn't Match", was correct in meaning but broke that symmetry — caught by rendering all three screens under Speculos and comparing them side by side, not by re-reading the string in isolation:

Invalid	Doesn't match	Matches
Invalid Secret / Recovery Phrase	Mismatched Secret / Recovery Phrase	Valid Secret / Recovery Phrase

Two limits worth stating plainly rather than glossing over:

• No test with a real, unbriefed user — the wording was reviewed by people who already know what it's supposed to mean, which is a weak check for whether it reads clearly cold.
• "Invalid" and "Mismatched" currently lead to the exact same next screen (check_result_callback(), unchanged by this PR) — the title tells the user why the phrase was rejected, but the app doesn't yet act differently on the two reasons. Out of scope here; noted for a later step.

Proof that BAGL is untouched

All six targets were built before and after this change (base = the tip of the ui_strings.h extraction this PR builds on, not bip85, to isolate this PR's own diff), with strings extracted from every .elf and diffed:

• nanos, nanox, nanos+: byte-for-byte identical — text/data/bss and sorted strings output, no difference at all.
• stax, flex, apex_p: diffs contain exactly the two new strings and the new titles array's debug symbol, nothing else.

Tests

tests/functional/test_bip39_seed_match.py extends from two verdict tests to three (its own docstring predicted this need), and its touch-side helper now asserts the title as well as the body line. One assertion needed a fix along the way: ragger's wait_for_text_on_screen() matches with re.match against each rendered line individually, anchored at that line's start — not a substring search over the whole screen. "not valid" failed because it isn't a line-start in "you have entered is not valid"; fixed to assert the full line. Found by querying Speculos's screen-content API directly rather than guessing.

tests/unit/tests/ui_strings.c's macro enumeration gets the two new macros (it doesn't track the header automatically — a renamed/removed macro is a compile error there, but a newly added one needs a matching line by hand, which this PR includes).

Verification

• Six unit-test-matrix configurations (ASan, UBSan, -O2, -Os, -fsigned-char, -funsigned-char): 80/80 each.
• Functional tests under Speculos, both stacks: nanox 9 passed / 9 skipped (was 8/9 before this PR's new test); stax 12 passed / 6 skipped (was 11/6). All skips are pre-existing device gating, unrelated to this change.
• clang-format --dry-run -Werror clean on touched src/ files.
• Not covered by this repository's CI on a fork PR: functional-test workflows don't run on pull requests from forks (read-only workflow permissions there), so the Speculos runs above are local and are the only evidence for this PR.