feat: parse ladder — three string-ladder rungs for weak-model drift (billion-context#1559)

[ranxianglei]

Three new rungs on the salvage ladder for weak-model syntax drift, per billion-context#1559 (filed during the v0.1.165 pre-release sweep). All changes confined to src/parse-compress-input.ts; no prompt/tool-surface change, no fixture churn, zero new dependencies.

Background

The 0.0.98 ladder (#470/#471) recovers object-in-slot and nested ranges, and parseObjectValue already routes non-JSON slot strings through the line-form path (splitLineEntries). Missing were the shapes where line-form text lives OUTSIDE any JSON shell:

• G1 — single-line slot string: {"content": "m00150-m00220 topic: summary here"} → was content-not-array, 0 ranges. Weak models / lossy gateways drop the line break of the taught header+body form; the text after the refs had nowhere to go (a single-line entry must have its body after a \n, so it always failed validation).
• G2 — top-level bare line-form string: compress("m00150–m00220 Auth\nsummary…") raw, or as a JSON string literal decoding to text → was malformed-json / not-object, 0 ranges. salvageContentEntries requires the "content": [ marker, so shell-less text had no recovery path at all.
• G3 — truncated stringified args whose content slot held a string: '"{"content": "m1–m5 A\nfull\nm6–m8 B\npar' → was truncated, 0 ranges. The array walker only handles [ slots, so every complete leading block was lost.

Changes

1. inlineSingleLineEntry (G1): when the line-form path yields exactly ONE part holding no line break, re-home the text after the refs header as the summary body by synthesizing the missing break. Applied to the bare-string paths only (object-slot branch + top-level salvage) — never to array elements, whose single-line header-only form has sibling-adoption semantics (#1001) and stays rejected.
2. salvageLineFormEntries + finishLineFormSalvage (G2): bare-string recovery with a new diagnostic flag stringSalvage: true (additive optional field on CompressParseDiagnostics). Adopted only when ≥1 valid range is found; otherwise the caller keeps its own verdict.
3. salvageStringSlotEntries (G3): extracts the surviving string-slot prefix (to the closing quote, or EOF when unterminated), decodes \n, returns the line-form blocks. Never-guess rule: while the slot is unterminated the trailing block may be mid-write and is dropped; an unterminated single block yields nothing. Reuses contentSalvage: true (set only on non-empty recovery), keeping parseCompressArgs: content given as a single entry object is rejected (content-not-array) while the same drift at the top level is tolerated — asymmetric salvage wall #470's flag semantics.
4. Diagnostic parity: the object-slot line-form branch's content-not-array rejection now carries invalidItems/invalidReasons (previously absent) so hosts can log WHY a shape was rejected.

Behavior changes (old → new)

Shape	Old	New
Single-line slot string, text after refs	content-not-array, 0 ranges	ok, 1 range (inline summary, derived topic, unflagged)
same, punctuation-only remainder ("m1–m5 …")	content-not-array, 0	unchanged (no letter/digit ⇒ not a summary)
Top-level bare line-form string (raw / quoted literal)	malformed-json / not-object, 0	ok, N ranges + stringSalvage: true
Top-level prose without refs	malformed-json / not-object	unchanged
Truncated stringified args, string slot, complete leading blocks	truncated, 0	truncated, N + contentSalvage: true
Truncated stringified args, string slot, single cut block	truncated, 0	unchanged (dropped, never guessed)
Canonical array/object forms, #470/#472 shapes	ok, flags as before	unchanged (pinned tests)
Array-element single-line header-only entry	no-valid-ranges	unchanged (#1001 adoption intact)

Design decisions

• Never split on embedded ref pairs in a single line. Prose citations (see m00510–m00520) are indistinguishable from headers; mis-splitting would silently corrupt a legitimate summary. A collapsed multi-block single line therefore recovers the FIRST block only (whole remainder becomes its summary) — under-recovery keeps the receipt honest so the model retries the rest. Pinned by test.
• G1 requires ≥1 letter/digit in the inline remainder (\p{L}|\p{N}), so "m1–m5 …" stays rejected.
• Salvage order for unparseable prefixes, most specific first: array objects → string slot → bare line-form. Existing paths keep precedence; nothing that recovered before loses recovery now.

Tests

15 new regression tests in tests/parse-compress-input.test.ts covering every row above, including the deliberate non-recovery pins. Full gate green on this head: typecheck ✓, 1217/1217 tests, build ✓, prettier stable ✓. Before/after verified with the A–K shape matrix from the issue thread.

Release plan

After merge: isolated release v0.0.99 PR (version bump only) → CI publish → verify npm view acp-kernel version → then the billion-context pin-bump PR referencing billion-context#1559.

中文摘要：为弱模型语法漂移补上解析梯级三档（单行槽内字符串内联摘要、顶层裸行式字符串回收、截断字符串槽完整块回收），全部带回归测试，1217/1217 全绿；合并后走 v0.0.99 发布流程，再升 billion-context 钉版本。

[github-actions]

📦 Built Package Artifact

Branch: 2026-09-28_string-ladder-rungs (68a7bb6)

Option A — Install from npm PR tag (recommended)

In your adapter project:

npm install acp-kernel@pr-477

Each push to this PR publishes a new version under the pr-477 npm tag.

Option B — Download artifact

1. Download the artifact from the Actions run
2. Extract the tarball and install:

tar xzf acp-kernel-pr477.tgz
npm install ./package

---

This comment is automatically updated on each push.

[ranxianglei]

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷 Review complete. I fetched the branch (c2b4dc7), ran the full gate locally on the head (Node v22.23.2: typecheck clean, tests 1217/1217, build success, format:check stable — matching CI, which is green: test 24/22, build-artifact, pr-validation), and independently re-derived your before/after table by running an identical 13-shape probe matrix against origin/master (5d7b538) and the PR head. Merge-base is the master tip, so no stale-base/rebase concern. Duplicate screen (open+closed issues, open PRs): no overlapping work. Findings below, in review-priority order.

① Direction (original issue) — billion-context#1559 asks for string-in-content-slot recovery in the salvage ladder. Verified against the issue: the approach matches its suggested fix (route through the existing line-form parser; contentSalvage reused for G3, new sibling stringSalvage for G2). One precision note: the issue's literal multi-line example ({"content": "m00150–m00220 topic\nsummary…"}) was already recovered on master by the pre-existing object-slot bare line-form path — my probe confirms ok n=1 on 5d7b538. The true residual gap was the single-line variant (no \n anywhere → one part → rejected "missing summary"), which is exactly what G1 targets. G2/G3 are siblings of the same root cause (line-form text outside any JSON shell), so bundling them here fits one-root-cause-one-fix. Direction: correct, fixes the root cause, not a symptom.

② Regression inventory — behaviors owned by the touched paths, each audited (probe = my A–M matrix; suite = full 1217):

1. Canonical object-array form → identical (ok n=1, unflagged) — probe-identical old/new.
2. parseCompressArgs: content given as a single entry object is rejected (content-not-array) while the same drift at the top level is tolerated — asymmetric salvage wall #470 object-in-slot salvage → identical (ok n=1 + contentSalvage).
3. Truncated canonical ARRAY slot salvage → identical (truncated n=1).
4. #1001 array-element header-only sibling adoption → identical (no-valid-ranges, 1 invalid) — G1 correctly never touches array elements.
5. Multi-line slot string (source-issue example) → identical (ok n=1, unflagged).
6. Top-level prose without refs (raw and quoted) → identical (malformed-json / not-object, unflagged) — no false-positive recovery.
7. Single cut block in truncated string slot → identical (truncated n=0, unflagged) — never-guess preserved.
8. Double-stringification unwrap precedence → unchanged (salvage order is most-specific-first: array objects → string slot → bare line-form; suite covers).
9. Single-quote repair retry → still functional and now less frequent (strictly-more-ranges win rule untouched; G3 single-quote test pins the interaction).
10. finishSalvage kind semantics → the hardcoded true for the salvage path is provably equivalent to the old looksTruncated(cleaned) (every salvage input is structurally truncated; verified against master source); empty-fallback still passes looksTruncated as before. Zero verdict drift for any previously-recovering input.

No unintended regressions found. Every changed line is additive recovery on inputs that previously yielded 0 ranges.

Behavior changes (old → new, per Deliverables rule 4):

• G1 single-line slot string: content-not-array/0 → ok/1, unflagged (deliberate — aligns with the existing unflagged multi-line slot-string lane).
• G2 bare top-level line-form / JSON string literal: malformed-json|not-object/0 → ok/N + stringSalvage:true.
• G3 truncated stringified args with string slot: truncated/0 → truncated/N + contentSalvage:true.
• Diagnostic parity (easy to miss): the object-slot line-form branch now populates diag.invalidItems/invalidReasons on the rejection path too (previously success-only). Verdict unchanged; e.g. {"content": "m1–m5 …"} goes inv=0 → inv=1 with the "missing summary" reason. Intentional and disclosed in the PR body; hosts reading these fields gain data, lose none.
• CompressParseDiagnostics.stringSalvage?: boolean — additive optional field.
• Judgment call I accept but flag: raw bare top-level text has no structural terminator signal, so a stream-cut bare string yields kind:"ok" with a possibly-trailing partial summary (refs intact — the load-bearing part). Consistent with "never guess, never fabricate"; just noting the asymmetry vs the shell-wrapped G3 lane.

③ Code — escape-aware scanner is correct (\\ skips pairs, unterminated detection, \n decode); anchored ref regexes carry no /g so .exec is stateless-safe; never-guest invariants hold at every rung (cut trailing block dropped, ≤1 block → nothing, punctuation-only remainder gated by /\p{L}|\p{N}/u). One soft observation, not a blocker: the source issue explicitly wanted the flag so "hosts can observe the drift rate", and G1 is unflagged — which means G1 drift becomes less observable than before (it used to show up as content-not-array rejections). My recommendation: leave it unflagged as-is for consistency with the closest precedent (multi-line slot strings, same slot, same payload type, differing only in line count — flagging only the single-line variant would make telemetry depend on incidental line count); if you later want real G1 drift-rate numbers, the right instrument is counting any string-in-slot regardless of line count, which is a separate semantic call. Say the word if you disagree and I'll fold a flag into the branch.

④ Tests — the 15 new tests are real regression tests: exact refs/summaries/kinds/flags asserted, and they pin the non-recovery cases (punctuation-only remainder, prose without refs, single cut block, #1001 array header-only) plus escape decoding and the quote-repair-retry interaction. Each rung regressing would flip an assertion. Valid.

⑤ Diff cleanliness — 2 files, +303/−9, every line serves the feature; no unrelated churn, no lockfile/generated artifacts, no mass whitespace. Clean.

Verdict: ready to merge (human action — I won't merge). Post-merge plan as proposed is sound: isolated release commit release v0.0.99 (version line + matching lockfile lines only) from master, then the billion-context pin-bump referencing #1559 — kernel-first ordering respected.

修复了弱模型漂移产生的三种字符串形态压测参数（单行槽内字符串、顶层裸字符串、截断字符串化参数）被整体拒绝的问题，在解析梯子上补了三档只救确定内容、绝不猜测的恢复；经本地全量门与 13 项前后行为矩阵对照确认无回归，仅有一处已披露的诊断字段增强，可以合并。