test(completion): cover the sentence helpers behind 33 surviving mutants

[acamarata]

Targeted from the mutation gate's actual surviving-mutant list for this file (run 34917215492): 33 missed, and all 33 sit in three private helpers.

function	missed mutants
find_sentence_start	18
find_sentence_end	13
the 200-char cap in extract_reasoning_sentence	2

Why 33 mutants lived despite 5 existing tests

The existing tests reached these helpers only indirectly, through parse_confidence, and asserted with contains(). Index arithmetic can be wrong by one in either direction and a contains() check still passes.

These call the helpers directly and assert exact indices. completion.rs drops 141 → 107 lines; the 5 pre-existing tests move verbatim into completion/tests.rs alongside 16 new ones.

Two of my own expected values were wrong — the code was right

On the first run, two assertions failed. I re-derived both from the algorithm independently rather than editing numbers until the test went green:

input	I wrote	correct	why
"one. two."	5	4	the trailing . has nothing after it so it is not a boundary; the inner ". " at index 3 wins
"One\nTwo. Three"	9	8	the ". " at index 7 is later than the newline at index 3

Both are now asserted at the correct value with the reasoning in a comment.

Verification — 14 applied, 14 killed

Including both index-arithmetic families the old contains()-style tests could never catch:

Mutant	Result
find_sentence_start -> 0 / -> 1	KILLED
while i > 0 → i < 0	KILLED
newline == → !=	KILLED
the ". " guard && → ||	KILLED
i + 1 < len → <=	KILLED
return i + 1 → i - 1	KILLED
find_sentence_end -> None	KILLED
Some(offset + i + 1) → - i / * i / i - 1	KILLED
terminator || → &&	KILLED
200-cap > → >= and > → ==	KILLED

The 200-cap pair needs boundary inputs specifically: a string of exactly 200 characters must come back whole (kills >=), and 201 must be truncated (kills ==).

cargo test --lib session::completion passes (21); clippy --all-targets --all-features -- -D warnings and cargo fmt --check both exit 0.