fix: keep soft-hyphen breaks at the insertion point

Fixes #162

Author: chenglou

AGENTS.md

@@ -99,7 +99,7 @@ See `DEVELOPMENT.md` for the current command surface and packaging/release check
 - URL-like runs such as `https://...` / `www...` are currently modeled as two breakable preprocessing units when a query exists: the path through the query introducer (`?`), then the query string. This is intentionally narrow and exists to stop obviously bad mid-path URL breaks without forcing the whole query string to fragment character-by-character.
 - Mixed app text also pulled in two more keep-worthy preprocessing rules: contextual escaped quote clusters like `\"word\"`, and numeric/time-range runs like `२४×७` / `7:00-9:00`.
 - For Southeast Asian scripts or mixed text containing Thai/Lao/Khmer/Myanmar, trust the `Range`-based corpus diagnostics over span-probing; span units can perturb line breaking there.
-- The remaining Chrome mixed-app `710px` soft-hyphen miss is extractor-sensitive and not cleanly local. Treat it as paragraph-scale / accumulation-sensitive until a cleaner reproducer appears, and do not patch the engine from only one extractor view.
+- The former Chrome mixed-app `710px` soft-hyphen miss is exact again after keeping chosen soft-hyphen breaks at the SHY boundary. Keep mixed app text as a regression canary, and do not reintroduce post-SHY grapheme packing from only one extractor view.
 - Safari `Range`-based probe extraction can over-advance across URL query text (`...path?q`) even when the real DOM height and the `span` extractor are exact. Cross-check `--method=span` before changing the engine on Safari URL/query probe misses.
 - Keep the current corpus lessons in mind:
   - Thai: contextual ASCII quotes were a real keep

CHANGELOG.md

@@ -11,6 +11,7 @@
 - CJK text followed by opening bracket annotations now wraps like browsers instead of leaving the opening bracket on the previous line (#148).
 - `{ wordBreak: 'keep-all' }` now handles no-space mixed Latin, numeric, and CJK text more like browsers.
 - Numeric prefix/postfix symbols like `$`, `%`, `€`, `+`, `−`, and `°` now stay attached to adjacent text the way browser line breaking does (#105).
+- Soft-hyphen breaks now stay at the soft-hyphen insertion point instead of pulling post-hyphen graphemes onto the broken line (#162).
 
 ## 0.0.5 - 2026-04-09
 

RESEARCH.md

@@ -15,7 +15,7 @@ This log is historical. The current practical steering picture is:
 - Myanmar still has two real canaries with residual Chrome/Safari disagreement around quote/follower-style classes, so it remains the main unresolved Southeast Asian frontier.
 - Urdu has a real Nastaliq/Naskh canary (`چغد`) with the same narrow-width negative field in Chrome and Safari, so it is clearly a shaping/context class rather than dirty data or a browser-only quirk. It remains parked rather than actively tuned.
 - Arabic coarse corpora are clean; the remaining work there is mostly a fine-width edge-fit class, not the old preprocessing/corpus-hygiene problems.
-- Mixed app text still matters because it catches product-shaped classes that books miss, especially soft-hyphen and extractor-sensitive cases.
+- Mixed app text is exact again in the maintained Chrome/Safari step10 sweeps and still matters as the product-shaped regression canary for URLs, emoji ZWJ runs, hard spaces, and soft hyphens.
 
 ## The problem: DOM measurement interleaving
 
@@ -220,8 +220,8 @@ Important keep:
 - model URL/query strings as narrow structured units, not one giant breakable blob
 
 Current status:
-- almost entirely clean
-- one remaining extractor-sensitive soft-hyphen miss around `710px` still looks paragraph-scale or accumulation-sensitive rather than like a neat local bug
+- exact in the maintained Chrome and Safari `step=10` sweeps
+- the former Chrome-only `710px` soft-hyphen miss went away after keeping chosen soft-hyphen breaks at the SHY boundary instead of packing post-SHY graphemes
 
 ### Thai
 

TODO.md

@@ -2,7 +2,7 @@ Current priorities:
 
 1. Keep the canaries honest
 
-- Mixed app text still has the extractor-sensitive `710px` soft-hyphen miss.
+- Mixed app text is exact again in the maintained Chrome/Safari step10 sweeps; keep it as a product-shaped regression canary rather than an active tuning target.
 - Chinese is still the clearest active CJK canary: Safari's step10 sweep is clean, while Chrome keeps a broader narrow-width positive field with real font sensitivity.
 - Myanmar and Urdu remain useful shaping/context canaries, but they are not the active tuning target right now.
 

accuracy/chrome.json

@@ -76825,5 +76825,5 @@
       "diff": 0
     }
   ],
-  "requestId": "1778228945426-cq4f58"
+  "requestId": "1778234152524-g3ze5g"
 }

accuracy/firefox.json

@@ -76825,5 +76825,5 @@
       "diff": 0
     }
   ],
-  "requestId": "1778228951223-53ai40"
+  "requestId": "1778234164610-4igphr"
 }

accuracy/letter-spacing.json

@@ -1,5 +1,5 @@
 {
-  "generatedAt": "2026-04-22T05:32:11.275Z",
+  "generatedAt": "2026-05-08T09:56:13.902Z",
   "browsers": [
     "chrome",
     "safari"
@@ -491,23 +491,23 @@
         "line": 1,
         "oursStart": 0,
         "browserStart": 0,
-        "oursEnd": 8,
+        "oursEnd": 6,
         "browserEnd": 7,
-        "oursText": "trans-at",
+        "oursText": "trans-",
         "browserText": "trans­a",
-        "oursRenderedText": "trans-at",
+        "oursRenderedText": "trans-",
         "browserRenderedText": "trans­a",
-        "oursContext": "trans­at|lantic transit",
+        "oursContext": "trans­|atlantic transit",
         "browserContext": "trans­a|tlantic transit",
-        "deltaText": "t",
+        "deltaText": "a",
         "reasonGuess": "our segment sum drifts from full-string width by 0.366px",
-        "oursSumWidth": 58.41593933105469,
-        "oursDomWidth": 58.0546875,
-        "oursFullWidth": 58.0499267578125,
+        "oursSumWidth": 43.52995300292969,
+        "oursDomWidth": 43.1640625,
+        "oursFullWidth": 43.1639404296875,
         "browserDomWidth": 46.609375,
         "browserFullWidth": 46.60194396972656
       },
-      "extractorSensitivity": null,
+      "extractorSensitivity": "range mismatch disappears with span",
       "message": null
     },
     {

accuracy/safari.json

@@ -76825,5 +76825,5 @@
       "diff": 0
     }
   ],
-  "requestId": "1778228947899-rtghd5"
+  "requestId": "1778234157599-qh89bv"
 }