feat(diff): hashline anchor comment localization (experimental, opt-in)

[adlternative]

What

Adds an opt-in, anchor-based comment localization path for code_comment, adapted from the hashline protocol (RimuruW/pi-hashline-edit).

When OCR_HASHLINE_ANCHORS=1 is set, the diff shown to the model renders every new-file line with a short anchor prefix:

42#KT:+	n, _ := strconv.Atoi(v)

The anchor hash is computed from the line and its two neighbors (xxh32(prev\0curr\0next) → 2 chars), so identical lines in different contexts get different anchors. The model localizes a comment by copying the anchor ("anchor": "42#KT" or a range "42#KT-45#MQ") instead of quoting a code snippet.

Why

The current existing_code text-matching pipeline resolves ambiguous matches by silently picking the first occurrence. Replaying real commit corpora through the production ResolveComment:

Localization method	opencode repo (95k added lines)	this repo (23k added lines)
existing_code, 1 line	63.2% correct	74.4% correct
existing_code, 3 lines	81.8% correct	93.0% correct
hashline anchor	100% correct	100% correct

41.5% of added lines in the TS corpus are not the first occurrence of their normalized content (imports, }, boilerplate) — exactly the silent-mislocation surface.

How

• internal/hashline (new): context hashing, strict ResolveSpec verification, AnnotateDiff.
• resolver.go: new highest-priority resolveFromAnchor step with two-factor validation — the hash is the checksum (both endpoints must verify), existing_code acts as a text hint that can veto hash collisions. On any mismatch it falls back to the existing hunk/file/relocation pipeline, so behavior is never worse than baseline. All resolution paths now record LlmComment.LocMethod (anchor / hunk / file / relocation / anchor_hint_veto) for observability.
• agent.go: env-gated annotation of {{diff}} and the injected diff map.
• examples/hashline_anchors/: anchor-aware tools.json + usage docs.

Cost / risk

• Anchor annotation adds +26% tokens on the diff text (tiktoken, p50 ≈ 24%). In a 5×5 end-to-end A/B on a live endpoint, total input tokens dropped ~25% — the model needed fewer file_read round-trips to confirm positions — and output tokens were flat.
• Anchor false-accept rate (model cites wrong line number, hash still verifies): ~0.5%, further reduced by the text-hint veto.
• Fully opt-in: default behavior is unchanged; a failed anchor degrades to the existing pipeline.

Testing

• go test ./... passes; new unit tests for hashing round-trips, collision behavior, diff annotation, and resolver anchor/veto/fallback paths.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

huzhening.hzn seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}