comisai 1.0.44 → 1.0.46

package/node_modules/@comis/core/dist/text/script-classes.js

@@ -4,23 +4,23 @@
  * classifier functions over it (classifyCodepointToRow, classifyCodepoint,
  * scriptShares, dominantScript).
  *
- * Single source of truth for per-script token factors (Phase 179 estimators),
- * FTS routing (Phase 180), and observability event classes (Phases 180/181).
+ * Single source of truth for per-script token factors (the context estimators),
+ * FTS routing, and observability event classes.
  * Adding a script later is a data edit — append a row, never a new mechanism.
  *
- * Defined in @comis/core so @comis/agent (Phase 179 estimators),
- * @comis/memory and @comis/skills (Phase 180 FTS routing), and the executor
- * (Phase 181 reply-language resolver) can import it without creating a
+ * Defined in @comis/core so @comis/agent (token estimators),
+ * @comis/memory and @comis/skills (FTS routing), and the executor
+ * (reply-language resolver) can import it without creating a
  * package cycle. NO imports from any @comis package — this file is pure
- * static data + pure functions, no I/O/clock/env (I9), and contains no
- * regex (V5 — zero ReDoS surface; classification is a single O(n)
+ * static data + pure functions, no I/O/clock/env, and contains no
+ * regex (zero ReDoS surface; classification is a single O(n)
  * codepoint-range scan, allocation bounded by the table size, never by input).
  * @module
  */
 /**
  * Per-script classification table. Scanned in declaration order
  * (first-match-wins): the hebrew/arabic MARKS rows precede their letter rows
- * so combining marks take the lower factor (I3). Ranges use numeric
+ * so combining marks take the lower factor. Ranges use numeric
  * `[lo, hi]` pairs (never regex) with per-range named comments, mirroring the
  * auditable boundary documentation of `hasCjkCodepoints` (lcd-fts.ts).
  * Every factor carries a measurement provenance comment; the only unmeasured
@@ -28,9 +28,10 @@
  */
 export const SCRIPT_CLASSES = [
     {
-        // hebrew MARKS — cantillation/te'amim + niqqud points (the FTS-02 strip set).
+        // hebrew MARKS — cantillation/te'amim + niqqud points (the
+        // normalizeForSearch strip set).
         // niqqud-bearing Hebrew measured 0.84 chars/token (46 chars → 55 tokens),
-        // qwen3-coder:30b probe 2026-06-12 (179-RESEARCH Pitfall 2); each mark
+        // qwen3-coder:30b probe 2026-06-12; each mark
         // ≈ 1 token → 0.1; refined by scripts/token-fixtures.
         class: "hebrew",
         ranges: [
@@ -44,11 +45,10 @@ export const SCRIPT_CLASSES = [
     {
         // hebrew letters — block + presentation forms.
         // unpointed chat Hebrew measured 2.71 chars/token → implied max 0.679
-        // (probe 2026-06-12); shipped 0.55. The TOK-02 corpus held for every
+        // (probe 2026-06-12). The token-fixtures corpus held for every
         // pure-Hebrew entry (worst implied 0.600) but the MIXED entry he_mixed_04
         // violated through the harmonic blend with latin LOCKED at 1.0 — implied
-        // letters bound 0.5016 — lowered by TOK-02 corpus, 2026-06-12
-        // (same-commit rule, plan 179-05).
+        // letters bound 0.5016, hence 0.5 (corpus measurement, 2026-06-12).
         class: "hebrew",
         ranges: [
             [0x0590, 0x05ff], // Hebrew block
@@ -59,8 +59,8 @@ export const SCRIPT_CLASSES = [
     {
         // arabic MARKS — harakat/tanween + superscript alef.
         // Same byte-level BPE behavior as Hebrew niqqud (≈ 1 token per mark) → 0.1;
-        // TOK-02 corpus measured the harakat entries at 1.26 chars/token aggregate
-        // (2026-06-12) — 0.1 holds; refined by scripts/token-fixtures.
+        // the token-fixtures corpus measured the harakat entries at 1.26 chars/token
+        // aggregate (2026-06-12) — 0.1 holds; refined by scripts/token-fixtures.
         class: "arabic",
         ranges: [
             [0x064b, 0x065f], // harakat + tanween
@@ -85,9 +85,9 @@ export const SCRIPT_CLASSES = [
     {
         // latin — A–Z, a–z, Latin-1 letters excl. × U+00D7 / ÷ U+00F7,
         // Extended-A/B, Extended Additional.
-        // TOK-02 corpus measured en at 5.16 chars/token aggregate, worst entry
-        // implied max 1.150 (2026-06-12) — 1.0 holds with margin.
-        // 1.0 LOCKED — I1 Latin byte-identity; never lower.
+        // The token-fixtures corpus measured en at 5.16 chars/token aggregate, worst
+        // entry implied max 1.150 (2026-06-12) — 1.0 holds with margin.
+        // 1.0 LOCKED — Latin byte-identity; never lower.
         class: "latin",
         ranges: [
             [0x0041, 0x005a], // A–Z
@@ -101,10 +101,10 @@ export const SCRIPT_CLASSES = [
     },
     {
         // cyrillic — single-sentence probe measured 3.32 chars/token → implied
-        // max 0.831 → shipped 0.75 (probe 2026-06-12). The TOK-02 corpus measured
-        // 13 ru chat/mixed violations (worst ru_chat_14 at 2.39 chars/token,
-        // implied max 0.598) — lowered by TOK-02 corpus, 2026-06-12
-        // (same-commit rule, plan 179-05).
+        // max 0.831 (probe 2026-06-12), but the token-fixtures corpus measured
+        // 13 ru chat/mixed violations against that bound (worst ru_chat_14 at
+        // 2.39 chars/token, implied max 0.598) — hence 0.59 (corpus measurement,
+        // 2026-06-12).
         class: "cyrillic",
         ranges: [
             [0x0400, 0x04ff], // Cyrillic
@@ -113,8 +113,8 @@ export const SCRIPT_CLASSES = [
         tokenFactor: 0.59,
     },
     {
-        // greek — measured 1.12 chars/token → implied max 0.279 (design's 0.8
-        // probe-disproven); 0.25 (probe 2026-06-12).
+        // greek — measured 1.12 chars/token → implied max 0.279;
+        // 0.25 (probe 2026-06-12).
         class: "greek",
         ranges: [
             [0x0370, 0x03ff], // Greek and Coptic
@@ -124,7 +124,7 @@ export const SCRIPT_CLASSES = [
     {
         // cjk — the six hasCjkCodepoints ranges PLUS the shipped trigger's two
         // gap ranges (CJK Symbols U+3000–303F, Ext-B U+20000–2A6DF). Yi (U+A000–)
-        // and Hangul Jamo (U+1100–) deliberately excluded (WR-01 over-match guard).
+        // and Hangul Jamo (U+1100–) deliberately excluded (over-match guard).
         // zh 1.73 / ja 1.36 chars/token → implied max 0.433/0.339; 0.3 covers both
         // (probe 2026-06-12).
         class: "cjk",
@@ -135,7 +135,7 @@ export const SCRIPT_CLASSES = [
             [0x3400, 0x4dbf], // CJK Extension A
             [0x4e00, 0x9fff], // CJK Unified Ideographs
             [0xac00, 0xd7af], // Hangul Syllables
-            [0xf900, 0xfaff], // CJK Compatibility Ideographs (NOT the literal glyph — WR-01)
+            [0xf900, 0xfaff], // CJK Compatibility Ideographs (numeric range, never a pasted glyph)
             [0x20000, 0x2a6df], // CJK Extension B (astral — gap range)
         ],
         tokenFactor: 0.3,
@@ -150,8 +150,8 @@ export const SCRIPT_CLASSES = [
         tokenFactor: 0.4,
     },
     {
-        // devanagari — measured 1.05 chars/token → implied max 0.261 (design's
-        // 0.5 probe-disproven); 0.25 (probe 2026-06-12).
+        // devanagari — measured 1.05 chars/token → implied max 0.261;
+        // 0.25 (probe 2026-06-12).
         class: "devanagari",
         ranges: [
             [0x0900, 0x097f], // Devanagari
@@ -160,7 +160,7 @@ export const SCRIPT_CLASSES = [
     },
     {
         // other — unmeasured — structurally unmeasurable (no single corpus exists
-        // for "everything else"); conservative 0.75 (design §4 TOK-02).
+        // for "everything else"); conservative 0.75.
         class: "other",
         ranges: [],
         tokenFactor: 0.75,
@@ -190,8 +190,8 @@ export function classifyCodepoint(cp) {
  * UTF-16-unit-weighted shares over non-neutral chars; the values sum to 1
  * when any non-neutral char exists; empty Map otherwise.
  *
- * Codepoint iteration (`for..of`) is surrogate-safe (Pitfall 9); weights are
- * UTF-16 units via `ch.length` (Pitfall 11) so shares line up with the
+ * Codepoint iteration (`for..of`) is surrogate-safe; weights are
+ * UTF-16 units via `ch.length` so shares line up with the
  * `.length` denominators the estimators divide.
  */
 export function scriptShares(text) {
@@ -213,11 +213,11 @@ export function scriptShares(text) {
     return shares;
 }
 /**
- * Non-Latin preference threshold for dominantScript. The pinned design
+ * Non-Latin preference threshold for dominantScript. The pinned
  * fixture ("ספר על docker") has 5 Hebrew vs 6 Latin units, so plain argmax
- * and strict-majority both return "latin" — WRONG (Pitfall 10): a Hebrew
- * utterance naming an English tool is Hebrew. OBS-01's mixed-code tolerance
- * (Phase 180) relies on code-dominated chunks (non-Latin share ≪ 0.30)
+ * and strict-majority both return "latin" — WRONG: a Hebrew
+ * utterance naming an English tool is Hebrew. The generation-quality
+ * mixed-code tolerance relies on code-dominated chunks (non-Latin share ≪ 0.30)
  * still returning "latin".
  */
 const DOMINANT_NON_LATIN_MIN_SHARE = 0.3;

package/node_modules/@comis/core/dist/text/slice-boundary.d.ts

@@ -1,24 +1,24 @@
 /**
- * Pure O(1) slice-boundary backoff (SAFE-01 / I9): no I/O/clock/env. Adjust a
+ * Pure O(1) slice-boundary backoff: no I/O/clock/env. Adjust a
  * raw UTF-16 cut index so a truncation never splits a surrogate pair or orphans
  * a base char's trailing combining/joiner/variation-selector sequence — the
- * result is mojibake-free for the scripts this milestone serves (Hebrew niqqud,
+ * result is mojibake-free for the served scripts (Hebrew niqqud,
  * emoji-ZWJ families, astral CJK).
  *
- * Single source of truth (I7): the content-length truncation cut sites in
+ * Single source of truth: the content-length truncation cut sites in
  * @comis/agent (tool-result-size-guard.ts) and @comis/daemon
  * (template-interpolation.ts) ALL import THIS symbol — never a copy. The helper
  * only MOVES an index backward; it injects nothing (so the truncation marker,
  * which stays plain English + newline-isolated, gains no bidi-control codepoint
- * at its boundary — I2).
+ * at its boundary).
  *
  * Detection uses Node's native ECMAScript Unicode property escapes under the `u`
  * flag (probed live on Node v22.21.1: `\p{M}`/`\p{Join_Control}`/
- * `\p{Variation_Selector}` all supported) — NOT `Intl.Segmenter` (REJECTED, I6:
+ * `\p{Variation_Selector}` all supported) — NOT `Intl.Segmenter` (REJECTED:
  * full grapheme segmentation of a 50K string to move two indices is O(n)
  * over-engineering for an O(1) need, and would add no dependency-free guarantee)
- * and NOT a reuse of the SCRIPT_CLASSES combining ranges (those are scoped to the
- * milestone's served scripts, not all Unicode marks). NO imports from any @comis
+ * and NOT a reuse of the SCRIPT_CLASSES combining ranges (those are scoped to
+ * the served-script set, not all Unicode marks). NO imports from any @comis
  * package — pure data + a pure function.
  * @module
  */
@@ -31,6 +31,6 @@
  *   truncation target).
  * @returns the adjusted index: never inside a surrogate pair, never immediately
  *   after a base char's trailing combining/joiner/VS run; bounded backoff, then
- *   cuts anyway. An ASCII boundary is a no-op (returns the index unchanged — I1).
+ *   cuts anyway. An ASCII boundary is a no-op (returns the index unchanged).
  */
 export declare function adjustSliceBoundary(text: string, index: number): number;

package/node_modules/@comis/core/dist/text/slice-boundary.js

@@ -1,30 +1,30 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Pure O(1) slice-boundary backoff (SAFE-01 / I9): no I/O/clock/env. Adjust a
+ * Pure O(1) slice-boundary backoff: no I/O/clock/env. Adjust a
  * raw UTF-16 cut index so a truncation never splits a surrogate pair or orphans
  * a base char's trailing combining/joiner/variation-selector sequence — the
- * result is mojibake-free for the scripts this milestone serves (Hebrew niqqud,
+ * result is mojibake-free for the served scripts (Hebrew niqqud,
  * emoji-ZWJ families, astral CJK).
  *
- * Single source of truth (I7): the content-length truncation cut sites in
+ * Single source of truth: the content-length truncation cut sites in
  * @comis/agent (tool-result-size-guard.ts) and @comis/daemon
  * (template-interpolation.ts) ALL import THIS symbol — never a copy. The helper
  * only MOVES an index backward; it injects nothing (so the truncation marker,
  * which stays plain English + newline-isolated, gains no bidi-control codepoint
- * at its boundary — I2).
+ * at its boundary).
  *
  * Detection uses Node's native ECMAScript Unicode property escapes under the `u`
  * flag (probed live on Node v22.21.1: `\p{M}`/`\p{Join_Control}`/
- * `\p{Variation_Selector}` all supported) — NOT `Intl.Segmenter` (REJECTED, I6:
+ * `\p{Variation_Selector}` all supported) — NOT `Intl.Segmenter` (REJECTED:
  * full grapheme segmentation of a 50K string to move two indices is O(n)
  * over-engineering for an O(1) need, and would add no dependency-free guarantee)
- * and NOT a reuse of the SCRIPT_CLASSES combining ranges (those are scoped to the
- * milestone's served scripts, not all Unicode marks). NO imports from any @comis
+ * and NOT a reuse of the SCRIPT_CLASSES combining ranges (those are scoped to
+ * the served-script set, not all Unicode marks). NO imports from any @comis
  * package — pure data + a pure function.
  * @module
  */
 /** Bounded backoff: a pathological combining-mark/joiner run cuts anyway after
- *  this many code units, so the helper stays O(1) on a 50K string (T-182-02). */
+ *  this many code units, so the helper stays O(1) on a 50K string. */
 const MAX_BACKOFF = 16;
 /** Trailing combiner/joiner/variation-selector test. `\p{M}` = Mn/Me combining
  *  marks; `\p{Join_Control}` = ZWJ (U+200D); `\p{Variation_Selector}` = the
@@ -40,7 +40,7 @@ const TRAILING_COMBINER = /\p{M}|\p{Join_Control}|\p{Variation_Selector}/u;
  *   truncation target).
  * @returns the adjusted index: never inside a surrogate pair, never immediately
  *   after a base char's trailing combining/joiner/VS run; bounded backoff, then
- *   cuts anyway. An ASCII boundary is a no-op (returns the index unchanged — I1).
+ *   cuts anyway. An ASCII boundary is a no-op (returns the index unchanged).
  */
 export function adjustSliceBoundary(text, index) {
     if (index <= 0)

package/node_modules/@comis/core/dist/text/token-factor.d.ts

@@ -1,14 +1,14 @@
 /**
  * Harmonic share-weighted token factor in (0, 1]; returns 1.0 for
- * empty/all-neutral/pure-ASCII text (I1 — Latin byte-identity). Shares are
+ * empty/all-neutral/pure-ASCII text (Latin byte-identity). Shares are
  * UTF-16 code units (matching the `.length` the estimators divide), weighted
  * per table ROW (not per class) so combining-mark rows carry their own,
  * lower factor than the letter rows of the same class.
  *
  * Why arithmetic mean is forbidden: tokens add per-class, so
  * `len/(R*f)` must equal `sum(len_i/(R*f_i))` — i.e. `1/f = sum(share_i/f_i)`
- * (harmonic). Probe evidence: the mixed he+latin design string measured 15
+ * (harmonic). Measured evidence: a mixed Hebrew+Latin string measured 15
  * qwen tokens — harmonic estimates 15 exactly, arithmetic estimates 14
- * (under-count) [179-RESEARCH Pattern 3, 2026-06-12].
+ * (under-count).
  */
 export declare function scriptTokenFactor(text: string): number;

package/node_modules/@comis/core/dist/text/token-factor.js

@@ -7,29 +7,29 @@
  *
  * Combination across script classes is HARMONIC per-row share summation
  * (`1/f = sum(share_i / f_i)`), NOT an arithmetic mean: tokens add per-class,
- * so `len/(R*f)` must equal `sum(len_i/(R*f_i))`. Probe evidence: a mixed
- * Hebrew+Latin string measured 15 real qwen tokens — harmonic estimates 15
- * exactly, arithmetic estimates 14 (an anti-conservative under-count)
- * [179-RESEARCH Pattern 3, qwen3-coder:30b probe 2026-06-12].
+ * so `len/(R*f)` must equal `sum(len_i/(R*f_i))`. Measured evidence: a mixed
+ * Hebrew+Latin string measured 15 real qwen tokens (qwen3-coder:30b) —
+ * harmonic estimates 15 exactly, arithmetic estimates 14 (an
+ * anti-conservative under-count).
  *
- * Pure function, same purity rules as script-classes.ts: no I/O/clock/env
- * (I9), no regex (V5). Imported relatively (same-package rule, never via the
+ * Pure function, same purity rules as script-classes.ts: no I/O/clock/env,
+ * no regex. Imported relatively (same-package rule, never via the
  * @comis/core barrel).
  * @module
  */
 import { classifyCodepointToRow } from "./script-classes.js";
 /**
  * Harmonic share-weighted token factor in (0, 1]; returns 1.0 for
- * empty/all-neutral/pure-ASCII text (I1 — Latin byte-identity). Shares are
+ * empty/all-neutral/pure-ASCII text (Latin byte-identity). Shares are
  * UTF-16 code units (matching the `.length` the estimators divide), weighted
  * per table ROW (not per class) so combining-mark rows carry their own,
  * lower factor than the letter rows of the same class.
  *
  * Why arithmetic mean is forbidden: tokens add per-class, so
  * `len/(R*f)` must equal `sum(len_i/(R*f_i))` — i.e. `1/f = sum(share_i/f_i)`
- * (harmonic). Probe evidence: the mixed he+latin design string measured 15
+ * (harmonic). Measured evidence: a mixed Hebrew+Latin string measured 15
  * qwen tokens — harmonic estimates 15 exactly, arithmetic estimates 14
- * (under-count) [179-RESEARCH Pattern 3, 2026-06-12].
+ * (under-count).
  */
 export function scriptTokenFactor(text) {
     // One O(n) pass accumulating per-ROW UTF-16-unit counts (rows, not classes —

package/node_modules/@comis/core/dist/text/trigram-query.d.ts

@@ -1,30 +1,30 @@
 /**
- * Script-routed FTS5 trigram query construction (FTS-01).
+ * Script-routed FTS5 trigram query construction.
  *
  * Routes a query to one of three lanes and, for the trigram lane, builds the
  * FTS5 MATCH string from untrusted user/model query text:
  *   - "word": NO non-Latin token → the caller uses the ORIGINAL string against
- *     the existing porter word lane (I1: byte-identical SQL for all-Latin).
+ *     the existing porter word lane (byte-identical SQL for all-Latin queries).
  *   - "tri":  at least one non-Latin token → a quoted, normalized, bounded
  *     MATCH expression over the trigram twins.
  *   - "scan": every non-operator token folds below the 3-codepoint trigram
  *     floor → the caller runs its bounded normalized scan floor over `scanTokens`.
  *
- * DECISION RECORD — Open Question 1 resolved as **Option B**: OR-of-trigram
- * decomposition ONLY for suffixing-morphology scripts (cyrillic + greek);
- * whole-quoted tokens for hebrew/arabic/cjk/everything else. A quoted whole
- * token is FTS5 substring semantics, which fails the pinned Russian criterion
- * (`книга` is not a substring of stored `книги`); an OR-of-trigrams group
- * matches it (RESEARCH probe 2, 2026-06-13). he/ar/CJK are prefix/substring
- * cases the whole-token shape already satisfies (the Hermes production
- * precedent). Adding a suffixing script later is a data edit to SUFFIXING_SCRIPTS.
+ * DECISION RECORD: OR-of-trigram decomposition ONLY for suffixing-morphology
+ * scripts (cyrillic + greek); whole-quoted tokens for hebrew/arabic/cjk/
+ * everything else. A quoted whole token is FTS5 substring semantics, which
+ * fails inflected-suffix recall (`книга` is not a substring of stored
+ * `книги`); an OR-of-trigrams group matches it. he/ar/CJK are prefix/substring
+ * cases the whole-token shape already satisfies. Adding a suffixing script
+ * later is a data edit to SUFFIXING_SCRIPTS.
  *
- * SECURITY (T-180-01-01 / T-180-01-02): every non-operator term is double-quoted
+ * SECURITY: every non-operator term is double-quoted
  * AFTER normalization, operators come only from the exact-uppercase allowlist,
  * any residual ASCII `"` is stripped from a term before quoting, and the only
  * parens are builder-emitted (never sourced from user text). MAX_QUERY_TOKENS +
  * MAX_TRIGRAMS_PER_TOKEN bound the MATCH size; the dangling-operator sweep
- * eliminates the probe-pinned FTS5 syntax-error shapes. Pure (I9): no I/O.
+ * eliminates the known FTS5 syntax-error shapes (a stranded leading/trailing
+ * or doubled operator is a hard FTS5 parse error). Pure: no I/O.
  * @module
  */
 /** The lane a query is routed to. */

package/node_modules/@comis/core/dist/text/trigram-query.js

@@ -1,39 +1,40 @@
 // SPDX-License-Identifier: Apache-2.0
 /**
- * Script-routed FTS5 trigram query construction (FTS-01).
+ * Script-routed FTS5 trigram query construction.
  *
  * Routes a query to one of three lanes and, for the trigram lane, builds the
  * FTS5 MATCH string from untrusted user/model query text:
  *   - "word": NO non-Latin token → the caller uses the ORIGINAL string against
- *     the existing porter word lane (I1: byte-identical SQL for all-Latin).
+ *     the existing porter word lane (byte-identical SQL for all-Latin queries).
  *   - "tri":  at least one non-Latin token → a quoted, normalized, bounded
  *     MATCH expression over the trigram twins.
  *   - "scan": every non-operator token folds below the 3-codepoint trigram
  *     floor → the caller runs its bounded normalized scan floor over `scanTokens`.
  *
- * DECISION RECORD — Open Question 1 resolved as **Option B**: OR-of-trigram
- * decomposition ONLY for suffixing-morphology scripts (cyrillic + greek);
- * whole-quoted tokens for hebrew/arabic/cjk/everything else. A quoted whole
- * token is FTS5 substring semantics, which fails the pinned Russian criterion
- * (`книга` is not a substring of stored `книги`); an OR-of-trigrams group
- * matches it (RESEARCH probe 2, 2026-06-13). he/ar/CJK are prefix/substring
- * cases the whole-token shape already satisfies (the Hermes production
- * precedent). Adding a suffixing script later is a data edit to SUFFIXING_SCRIPTS.
+ * DECISION RECORD: OR-of-trigram decomposition ONLY for suffixing-morphology
+ * scripts (cyrillic + greek); whole-quoted tokens for hebrew/arabic/cjk/
+ * everything else. A quoted whole token is FTS5 substring semantics, which
+ * fails inflected-suffix recall (`книга` is not a substring of stored
+ * `книги`); an OR-of-trigrams group matches it. he/ar/CJK are prefix/substring
+ * cases the whole-token shape already satisfies. Adding a suffixing script
+ * later is a data edit to SUFFIXING_SCRIPTS.
  *
- * SECURITY (T-180-01-01 / T-180-01-02): every non-operator term is double-quoted
+ * SECURITY: every non-operator term is double-quoted
  * AFTER normalization, operators come only from the exact-uppercase allowlist,
  * any residual ASCII `"` is stripped from a term before quoting, and the only
  * parens are builder-emitted (never sourced from user text). MAX_QUERY_TOKENS +
  * MAX_TRIGRAMS_PER_TOKEN bound the MATCH size; the dangling-operator sweep
- * eliminates the probe-pinned FTS5 syntax-error shapes. Pure (I9): no I/O.
+ * eliminates the known FTS5 syntax-error shapes (a stranded leading/trailing
+ * or doubled operator is a hard FTS5 parse error). Pure: no I/O.
  * @module
  */
 import { classifyCodepoint } from "./script-classes.js";
 import { normalizeForSearch } from "./normalize-search.js";
 /**
  * DoS bound on the token count fed into the MATCH builder. The upstream FTS5
- * sanitizer does NOT cap token count, so the builder must (RESEARCH Security
- * Domain — "DoS via pathological queries"). 16 is generous for real queries.
+ * sanitizer does NOT cap token count, so the builder must — a pathological
+ * query could otherwise inflate the MATCH string without limit. 16 is
+ * generous for real queries.
  */
 const MAX_QUERY_TOKENS = 16;
 /**
@@ -45,7 +46,7 @@ const MAX_TRIGRAMS_PER_TOKEN = 12;
 /**
  * Scripts whose morphology is SUFFIXING — a stem keeps its prefix but changes
  * its tail across inflections, so substring (quoted whole-token) matching fails
- * (`книга`/`книги`). These get OR-of-trigram decomposition (OQ-1 Option B).
+ * (`книга`/`книги`). These get OR-of-trigram decomposition.
  * Everything else (hebrew/arabic/cjk/...) is prefix/substring morphology and
  * stays whole-quoted. Adding a script here is a one-line data edit.
  */
@@ -93,7 +94,7 @@ function orOfTrigrams(normalized) {
     return `(${trigrams.join(" OR ")})`;
 }
 /** Double-quote a term for FTS5. The caller has already stripped any interior
- *  ASCII `"` (so the quotes can never be unbalanced — T-180-01-01). */
+ *  ASCII `"` (so the quotes can never be unbalanced). */
 function quote(term) {
     return DQUOTE + term + DQUOTE;
 }
@@ -115,7 +116,7 @@ function stripQuotes(s) {
  * cross-package import between core and skills is forbidden in both directions,
  * so the source is duplicated with this citation rather than imported. These are
  * fixed alternations over a 3-word operator set: no nested quantifiers, no
- * backtracking surface (T-180-01-02).
+ * backtracking surface (ReDoS-safe).
  */
 function sweepDanglingOperators(match) {
     let working = match;
@@ -179,7 +180,8 @@ export function routeSearchQuery(query, opts) {
     const tokens = tokenize(raw).slice(0, MAX_QUERY_TOKENS);
     if (tokens.length === 0)
         return { lane: "word" };
-    // Step 3: all-Latin → word lane (caller uses the ORIGINAL string — I1).
+    // Step 3: all-Latin → word lane (caller uses the ORIGINAL string, so the
+    // existing porter word lane emits byte-identical SQL).
     const anyNonLatin = tokens.some((t) => hasNonLatin(t.text));
     if (!anyNonLatin)
         return { lane: "word" };
@@ -213,10 +215,11 @@ export function routeSearchQuery(query, opts) {
         }
         scanTokens.push(normalized);
         // Drop tokens below the 3-codepoint trigram floor — correctness-critical:
-        // a <3-cp token in an AND context returns ZERO ROWS (RESEARCH probe 3).
+        // a <3-cp token can never match a trigram index, so in an AND context it
+        // returns ZERO ROWS for the whole query.
         if (codepointLength(normalized) < 3)
             continue;
-        // OQ-1 Option B: suffixing-script token (≥4 cp) NOT directly governed by a
+        // Suffixing-script token (≥4 cp) NOT directly governed by a
         // preceding NOT → OR-of-trigrams group; else whole-quoted.
         if (hasSuffixingScript(normalized) && codepointLength(normalized) >= 4 && !governedByNot) {
             terms.push(orOfTrigrams(normalized));

package/node_modules/@comis/core/dist/tool-metadata.d.ts

@@ -77,38 +77,38 @@ export interface ComisToolMetadata {
     mcpExportPolicy?: "safe" | "permission-gated" | "never-export";
     /** When true, this tool's lifecycle produces no activity messages (the
      *  activity pipeline skips it). Lifecycle reactions + final delivery are
-     *  unaffected. Read by the activity layer (§16.11). */
+     *  unaffected. Read by the activity layer. */
     suppressActivity?: boolean;
     /** When true, a non-zero `details.exitCode` in this tool's RESULT is the DRIVEN
      *  subject's exit code (informational), NOT the tool's own outcome — so the bridge's
-     *  exit-code failure heuristic (§D1) must NOT flag the call as failed. Set on the
+     *  exit-code failure heuristic must NOT flag the call as failed. Set on the
      *  terminal-driver perception tools (status/read/wait), whose result reports the driven
      *  session's `exitCode`. Without it a driven program exiting non-zero (e.g. bash `exit 1`)
-     *  misclassifies a perfectly-successful `terminal_session_status` as a tool failure
-     *  (real-VPS 2026-06-16). Default (absent) ⇒ the heuristic applies (exec/process, where
+     *  misclassifies a perfectly-successful `terminal_session_status` as a tool
+     *  failure. Default (absent) ⇒ the heuristic applies (exec/process, where
      *  the exit code IS the tool's outcome). */
     exitCodeIsDrivenSession?: boolean;
     /** Tool-specific failure classifier consulted *before* the `tool:executed`
-     *  emit (§16.10), so observability never sees the raw result. Receives the
+     *  emit, so observability never sees the raw result. Receives the
      *  tool result and the SDK `isError` flag; returns `true`/`false` (failed or
      *  not) or `{ errorKind, … }` (failed, with a closed-union classification plus
      *  optional verdict provenance). Lets a tool flag a logically-failed result
-     *  that the SDK reported as success (e.g. a non-zero exit code). (§16.11).
+     *  that the SDK reported as success (e.g. a non-zero exit code).
      *
-     *  Verdict provenance (P2/D2) — all optional, additive on the contract:
+     *  Verdict provenance — all optional, additive on the contract:
      *  - `classifiedField`: which STRUCTURED field drove the verdict (never the body).
      *  - `matchedRule`: the literal regex/rule description that matched (a fixed
      *    string, not a serialized RegExp and not tool-output data).
      *  - `matchedToken`: the concrete token that matched (e.g. a status code). This
      *    is the only provenance field that may carry tool output, so downstream
-     *    sinks (Plan 04) bound it via `sanitizeLogString(...).slice(0, 1500)`. */
+     *    log sinks bound it via `sanitizeLogString(...).slice(0, 1500)`. */
     failureDetector?: (result: unknown, isError: boolean) => boolean | {
         errorKind: ErrorKind;
-        /** Which structured field drove the verdict (P2/D2). */
+        /** Which structured field drove the verdict. */
         classifiedField?: "error" | "status" | "message" | "failures";
-        /** The regex/rule literal that matched (P2/D2). */
+        /** The regex/rule literal that matched. */
         matchedRule?: string;
-        /** The concrete token that matched, e.g. a status code (P2/D2). */
+        /** The concrete token that matched, e.g. a status code. */
         matchedToken?: string;
     };
 }

package/node_modules/@comis/core/dist/workspace/templates.d.ts

@@ -10,9 +10,9 @@ export type WorkspaceFileName = (typeof WORKSPACE_FILE_NAMES)[number];
  * Platform-owned workspace files — content-hash-refreshed by `ensureWorkspace`
  * on every invocation when the on-disk sha256 differs from `DEFAULT_TEMPLATES`.
  *
- * 2026-05-20 lineage: the prior `wx`-only seed flow let stale pre-fix templates
- * persist forever, so a sub-agent's inherited workspace kept promising a
- * pre-warmed venv that no code provisioned. These three files declare themselves
+ * A `wx`-only seed flow would let stale templates persist forever, so a
+ * sub-agent's inherited workspace could keep promising behavior no code
+ * provisions. These three files declare themselves
  * read-only platform contract in their own prose — this partition makes that
  * contract real.
  */

package/node_modules/@comis/core/dist/workspace/templates.js

@@ -20,9 +20,9 @@ export const WORKSPACE_FILE_NAMES = [
  * Platform-owned workspace files — content-hash-refreshed by `ensureWorkspace`
  * on every invocation when the on-disk sha256 differs from `DEFAULT_TEMPLATES`.
  *
- * 2026-05-20 lineage: the prior `wx`-only seed flow let stale pre-fix templates
- * persist forever, so a sub-agent's inherited workspace kept promising a
- * pre-warmed venv that no code provisioned. These three files declare themselves
+ * A `wx`-only seed flow would let stale templates persist forever, so a
+ * sub-agent's inherited workspace could keep promising behavior no code
+ * provisions. These three files declare themselves
  * read-only platform contract in their own prose — this partition makes that
  * contract real.
  */

package/node_modules/@comis/core/dist/workspace/workspace-manager.d.ts

@@ -38,8 +38,8 @@ export interface EnsureWorkspaceOptions {
      * `write` tool can overwrite the seed without tripping the read-before-write
      * (`[not_read]`) gate.
      *
-     * Left undefined at daemon startup (no session yet) — the seeded files will
-     * be registered lazily on first read, same as before.
+     * Left undefined at daemon startup (no session yet) — the seeded files are
+     * registered lazily on first read.
      */
     tracker?: WorkspaceSeedTracker;
 }
@@ -87,9 +87,8 @@ export interface RegisterWorkspaceResult {
  * (before any session tracker exists) and seeds the agent's own workspace
  * files. Every subsequent session's first `write` to those paths would
  * otherwise hit a read-before-write gate. Call this right after the
- * per-session tracker is created (or per-turn, before the session-lifetime
- * registry existed) so the agent's first `write` to its own workspace
- * passes the gate.
+ * per-session tracker is created so the agent's first `write` to its own
+ * workspace passes the gate.
  *
  * Idempotency: when the tracker already records a matching-mtime entry
  * for a path (via the optional `getReadState` method), the helper skips