fix(lm): paragraph is a hard prediction boundary

[willwade]

What this fixes

Dasher-Windows tester report (v0.1.33): "When you zoom an "enter", the canvas is supposed to reset right? But it doesn'"'"'t, it remembers the last punctuation you put. Then you have to zoom a capital letter."

Root cause

The PPM context chained through the paragraph symbol. The bundled English corpus breaks paragraphs as .\r\n (94% of lines start with a capital after CRLF), so the [.][para] context matched training and predicted sentence start — but any other predecessor of the newline produced a sequence the corpus never contains, and the escape mixture degraded the after-newline prediction to a mid-sentence blend:

context before newline	uppercase-group mass (before → after)
. + newline	0.95 → 0.92
bare word + newline	0.67 → 0.92
zoomed space + newline	0.67 → 0.92
comma + newline	0.72 → 0.92
blank line (double newline)	0.67 → 0.92

(Measured with dasher_get_probabilities against the real engine + bundled corpus; "uppercase-group mass" = the share of the top-level Upper case Latin letters group in the crosshair distribution.)

Fix

A paragraph now starts the canonical paragraph-start context: the alphabet default context (trailing whitespace stripped — the corpus has no [space][para] sequences) followed by the paragraph symbol. Three sites, all in CAlphabetManager:

• GetContextSymbols (seeded/rebuilt contexts): symbols before the last paragraph symbol are replaced by the canonical seed symbols
• CreateSymbolNode (live zoom): the paragraph node'"'"'s context is the canonical seed context rather than the parent chain extended
• CSymbolNode::SetFlag NF_COMMITTED (adaptive learn): the trie keeps the learned [x][para] chain, but the node'"'"'s prediction context becomes the canonical seed

Training is unchanged — PPM vine learning already records after-paragraph statistics at every order. Alphabets without a paragraph symbol are untouched (SymbolPrintsNewLineCharacter guards every site).

Verification

• Full suite green: 46/46 incl. PPM goldens, LM correctness, direct-mode shadow, property invariants
• New mapping-free regression test (test_paragraph_context.cpp): ten predecessor variants must yield exactly the canonical distribution, and it must differ from mid-word / bare-period contexts

Follow-up needed on the Dasher-Windows side after this releases: bump the submodule pin.

Confidence Score: 5/5

The PR appears safe to merge; the remaining feedback is a non-blocking test gap.

Findings

1.  Live newline paths lack coverage ▶

Summary

The PR makes newline predictions start from a consistent paragraph context.

• Paragraph breaks now reset Dasher’s next-character predictions.

No new actionable issues were found.

_{Reviews (5) · Last reviewed commit: "fix(lm): paragraph is a hard prediction ..." · Reviewed by Greptile}