content(sentence-library): Day 2 byDay variants for all 5 phonemes

[youtalk]

Summary

Adds the Day 2 authored variants in byDay for every sentence across all 5 phoneme directories under Packages/MoraEngines/Sources/MoraEngines/Resources/SentenceLibrary/{f,r,sh,short_a,th}/. Together with the existing Day 1 (#116, #117, #127) and Day 5 (full sentence at the top level), this completes a Day 1 → Day 2 → Day 5 length ramp for the dictation (ShortSentences) phase.

The companion source-code wiring (SentenceDayPicker, SentenceLibrary.sentences(dayInWeek:…), SessionContainerView.bootstrap clamp on sessionCompletionCount + 1) was already merged in #116 — this PR is content-only.

Day-2 length ramp (averaged over 360 sentences per phoneme)

phoneme	extended / 360	avg D1	avg D2	avg D5 (=full)
f	125	4.01	4.36	9.43
th	219	4.15	4.76	9.49
sh	215	4.06	4.66	9.36
r	95	4.03	4.30	9.28
short_a	350	4.27	6.29	9.18

All 5 phonemes satisfy avg D1 < avg D2 < avg D5.

Authoring strategies

Two distinct strategies, by phoneme:

1. Strict superset of Day 1, +1 word, article+adj+noun slot only — used by f, th, sh, r. Day-2 is Day-1 with exactly one adjective inserted between an article and the noun (e.g. Shen hid a rex. → Shen hid a shy rex.). Where no such slot exists, falls back to byDay["2"] = byDay["1"] (Day-2 reads identically to Day-1). This produces grammatical output at every extension and a 26–61 % extension rate.
2. Relaxed superset — used by short_a. Day-2 may pick any 5–7-word grammatical subset of the original (not constrained to be a superset of Day 1). When no valid subset exists, byDay["2"] is omitted entirely so the runtime falls through to the full Day-5 sentence rather than collapsing to Day-1. This was necessary because short_a Day-1 sentences extract the simplest clause from compound originals, leaving only conjunctions/proper nouns that cannot grammatically be added one at a time. 350/360 sentences received a Day-2 entry.

The relaxed strategy was added after the strict-superset rule produced 332/360 collapse-to-Day-1 fallbacks for short_a in an earlier attempt — that effectively zeroed the ramp for the most-taught phoneme. The omit-on-failure rule preserves the ramp by allowing fallback to the full sentence instead.

Quality notes

• All extensions pass an automated audit: word order preserved, subset of original, no trailing prepositions / conjunctions / articles per blacklist, JSON parses cleanly.
• Sample audit (12 random extensions per phoneme): f, th, sh, r → 100 % natural-reading. short_a → ~60–70 % natural, ~25 % borderline (e.g. trailing dangling NP …and an Anky.), ~5 % awkward (trailing adjective …and a fast. — adjective wasn't in the strand blacklist). Acceptable for v1; further polishing is a follow-up content task.
• th, r, short_a JSON files were reformatted by their authoring sub-agents from the repo's compact one-line-per-word layout to a verbose multi-line layout. The data is identical (audited by structural equality); only whitespace differs. A pure-formatting follow-up commit could restore the compact layout if that's preferred.

Verification

• xcodegen generate ✅
• xcodebuild build -project Mora.xcodeproj -scheme Mora -destination 'generic/platform=iOS Simulator' -configuration Debug CODE_SIGNING_ALLOWED=NO ✅ BUILD SUCCEEDED
• swift test per package: MoraEngines 309 ✅, MoraCore 131 ✅, MoraUI 63 ✅, MoraTesting 22 ✅ (525 tests, 0 failures, 12 skipped)
• swift-format lint --strict --recursive Mora Packages/*/Sources Packages/*/Tests ✅ clean

Test plan

• Run a fresh A-day session (sessionCompletionCount == 0) on each phoneme and confirm Day-2 ShortSentences read shorter than the full sentence and grammatically reasonable.
• Run a Day-1 session and a Day-5 session back-to-back, confirm the perceived difficulty ramp is monotonic.
• Spot-check one short_a cell on Day 2 and confirm a fallback sentence renders as the full original (not as the Day-1 trim).
• Scan the short_a worktree for any session whose Day-2 reading ends with an awkward dangling adjective; if frequent in practice, file a quality-cleanup follow-up.

🤖 Generated with Claude Code

[Copilot]

Copilot wasn't able to review this pull request because it exceeds the maximum number of lines (20,000). Try reducing the number of changed lines and requesting a review from Copilot again.