docs: drop finished work from TODO.md and rehome the PIT exclusion rationale

[bernardladenthin]

Summary

• TODO.md's own header said completed work belongs in git history and crossrepostatus.md, not in the file — and the longest entry in it opened by declaring itself finished ("Still out … nothing, as it turns out", 33 lines). It is deleted; the file goes from 102 to 51 lines with 7 genuinely open bullets.
• What that entry carried is not history, and it is kept: the reason two classes can never reach 100 % mutation coverage. Compressed and moved into CLAUDE.md's PIT section, where guidance for working in this repo belongs.

Why that rationale had to survive

Both classes' survivors are equivalent mutants — unkillable through the public API, not merely untested. Without the write-up the next person re-attempts an impossible target:

• document.AiMdHeaderCodec — the colon guard in read is reached only after startsWith(HEADER_FIELD_PREFIX), so colonIndex is either -1 or >= 2. The < 0 boundary mutant differs only at the unreachable 0; the +1 → -1 mutant only at an empty field key, which values.put("", value) swallows because no header field is keyed "".
• support.AiSourceChunker (28/34) — its three observable boundaries are pinned (maxChars == 1, the end < length guard, the lastNewline > pos guard), but six mutants have no observable effect at all: the ArrayList capacity hint, return chunks vs emptyList() on empty input, two clamps the surrounding Math.max absorbs, the loop head the end >= length break already guarantees, and select's total <= maxChunks, where the equality branch computes the identity mapping.

Reaching 100 % on either would mean exposing internals or deleting a deliberately defensive guard.

The @VisibleForTesting entry was misleading and is sharpened

It claimed "No usages currently in any module", which reads as "nothing to do". The audit is about members that exist and are unannotated, and there are ten: four on LlamaCppJniAiGenerationProvider plus its two static mappers, two on AiMdDocumentCodec, AiPromptPreparationSupport.trimSourceAtLineBreak, and the three mojos' build*Configuration(). It now names them and states the decision it needs — a project-local marker annotation (support.ConvertToRecord is precedent) versus recording that the convention is not adopted here — so it can be closed rather than re-audited forever.

Test plan

• Diff is TODO.md + CLAUDE.md only — no source, test or build file touched
• Rebased onto current main (which has since taken feat!: send enable_thinking only when it was actually set #199 and fix(build): make srcmorph/pom.xml parseable again #200); verified the file no longer mentions enable_thinking, flashAttn or the PIT-scope item, all three of which are now closed
• CI is green on this branch — no: main pins the unpublished net.ladenthin:llama:5.2.0, so every srcmorph PR fails at dependency resolution until that release lands. Unrelated to this change.

Related issues / PRs

Part of a four-repo pass; siblings are the matching claude/todo-cleanup branches in java-llama.cpp, BitcoinAddressFinder and streambuffer. Each file's header now states the rule outright: finished items are deleted, not annotated.

Checklist

• I have read CONTRIBUTING.md and CODE_OF_CONDUCT.md
• My commits follow Conventional Commits
• No security-sensitive changes

---

Generated by Claude Code

[claude]

Review Summary

This is a clean documentation-only change that appropriately consolidates finished work and preserves important technical guidance.

What's Good ✓

1. Correct application of stated principles — TODO.md's own header says completed work belongs in git history and crossrepostatus.md, not in the file. Deleting the 33-line finished entry follows this correctly.

2. Important technical knowledge preserved — The rationale for why AiMdHeaderCodec and AiSourceChunker cannot reach 100% mutation coverage is genuinely valuable: both have equivalent mutants (unkillable through the public API, not merely untested). This prevents future work from re-attempting an impossible task. The explanations are technically sound:

   • AiMdHeaderCodec (line 164): After startsWith("- "), colonIndex is -1 or ≥2. Mutating < 0 to <= 0 only affects the unreachable 0; mutating +1 to -1 only affects empty-key case, which values.put("", value) swallows silently. ✓
   • AiSourceChunker (28/34): Six unobservable mutants correctly identified—capacity hint, empty-return equivalence, clamps absorbed by Math, loop conditions already covered, identity mapping in select equality branch. ✓

3. Improved header clarity — New phrasing makes the rule explicit: "Completed work is not recorded here... a finished item is deleted rather than annotated."

4. Actionable refinement — The @VisibleForTesting entry now lists specific members (4 on LlamaCppJniAiGenerationProvider + 2 static, 2 on AiMdDocumentCodec, AiPromptPreparationSupport.trimSourceAtLineBreak, 3 mojos) and frames the decision clearly, replacing the misleading "No usages currently."

No Issues Found

• No correctness problems
• No security implications
• Placement in CLAUDE.md PIT section is appropriate
• Diff touches only documentation as stated

Looks good to merge. ✓