v0.3.35

[0.3.35] — 2026-06-23

Fixes patch_source over-strict old_string matching: the TrioBASIC compiler is case-insensitive, but the matcher used StringComparison.Ordinal (case-sensitive), so when the AI remembered identifiers in all-lower/upper case, old_string not found was returned and it retried repeatedly. Also fixes an offset bug in the TrimEnd fuzzy branch.

Fixed

• patch_source matching is now case-insensitive (TrioBASIC semantics) — FindText previously used IndexOf(Ordinal) for strict comparison, while the TrioBASIC compiler treats cVR_x / cvr_x / CVR_X as identical. After reading source, the AI rewrites old_string from semantic memory and frequently lowercases mixed-case identifiers (e.g. cVR_bakdecelper → cvr_bakdecelper) — legal and equivalent in TrioBASIC, but rejected as "different strings" by the case-sensitive matcher → old_string not found → repeated retries (see chat_history/20260621_202830.json msg#68-70: 3 consecutive failures). FindText now adds an OrdinalIgnoreCase fallback after the exact match, aligning with the target language's semantics. CountOccurrences (the uniqueness check) is switched to OrdinalIgnoreCase as well, eliminating the "match found but count is 0" semantic split.
• Fixed offset bug in FindText's TrimEnd branch — The old source.IndexOf(srcLines[i], source.Length - source.Length) had length - length always equal to 0, equivalent to "find the line text's first occurrence from the file start". When old_string was a multi-line block and a line inside it recurred elsewhere in the file, the returned offset pointed to a different identical line → replacement at the wrong location. Now computes the true offset by summing line lengths.
• FindText returns a (pos, length) tuple; replacement slices by the actual match length — In the Trim fuzzy branch, source lines may carry trailing/leading whitespace, making the real matched block longer than old_string. The old code sliced with op.oldString.Length, cutting in the middle of the block and leaving a half-residue. Now returns the block's true length and the caller slices accordingly.

Behavioral impact

• AI mis-remembering identifier case → previously always failed and retried; now matches and replaces directly.
• Exact match still takes priority (when case is correct the fallback layer is not reached — behavior unchanged).
• Uniqueness protection is NOT relaxed: fragments that recur (e.g. the VR(cvr_hmibtn0)=0 clear-to-zero pattern across multiple subroutines) still require the AI to supply more context to locate — this remains a reasonable safety guard.