Replace clustering with a catalogue, and hand-write the references to judge it against

The pipeline's embed-and-cluster step is dead, and this commit holds both the
evidence for that and the step proposed to replace it.

Predicaments. Scenes are re-described as "what the person is up against", with
no names, jobs or places, then embedded and clustered (redescribe.py,
topic_words.py, topic_share.py). The pilot chose qwen3:14b over 3b by reading
both side by side. Two defects the pilot exposed are fixed: split.py missed
titles in quotes and a contents subtitle after a dash, so three stories had been
merged into their neighbours, and strip_names.py read New York place names as
people. The corrected corpus is probe/v2 (97 stories, 839 scenes);
carry_summaries.py reuses the 829 unchanged v1 summaries. Topic share fell from
20% to 13% at k=60, short of the pre-registered 10%.

Hand references. Three corpora were read scene by scene and written up by hand,
under the same prompt rules the local models get, as a baseline to judge them
against: O. Henry (probe/v2/claude, 839 scenes, 20 situations), Wharton's
Descent of Man (probe/wharton, 262 scenes, 16 groups) and Jacobs's The Lady of
the Barge (probe/jacobs, 157 scenes, 19 groups). Each has its own README and a
readable page. No inference was used for any of them.

Catalogue. probe/catalogue maps every hand group in the three references onto 36
situation entries, with an answer key per corpus and one recurrence rule applied
to all three. classify.py assigns a scene one entry or none, leave-one-corpus-
out; score.py checks it against the key, with a self-test on random labels.

Why clustering is out: hand-written predicaments, embedded and clustered exactly
as the model's were, agree with the hand grouping at ARI 0.05 — no better than
the 14B text's 0.07. Better rewriting cannot rescue it. Embeddings cannot even
shortlist: the hand label is the nearest entry 13% of the time and in the top 8
half the time.

The classification runs are not here. The dev and test runs are pre-registered
in probe/catalogue/README.md with the bar set beforehand, and are blocked on the
inference host, whose GPU has fallen off the PCIe bus three times. The 30-scene
partial output in out/ is not a result.

Review page. The situation review is now a browser page rather than JSON edited
by hand (review_page.py, review_page_logic.cjs with Node tests, format schema
v2). It has never been rendered in a real browser.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014BygvsUXV9eU6oHkTCkKZ1

tools/story-to-pack/probe/v2/gpu-pipeline.log

@@ -0,0 +1,111 @@
+18:34:43
+== carry v1 summaries, summarise new scenes (qwen2.5:3b-instruct)
+  new scene 29 (DOUGHERTY'S EYE-OPENER): Patrick, the Honorable, attempts to seduce Mrs. Dougherty but is thwarted by Jimmy, who responds aggressively.
+  new scene 30 (LITTLE SPECK IN GARNERED FRUIT): The newlyweds, newly arrived in their lavish flat, debate their honeymoon plans while the groom prepares to leave for a boxing match.
+  new scene 31 (LITTLE SPECK IN GARNERED FRUIT): The Kid, seeking a peach for the bride's favor, confronts multiple fruit vendors and a reluctant Justus O'Callahan, each unwilling to part with the desired fruit before its season.
+  new scene 32 (LITTLE SPECK IN GARNERED FRUIT): The Kid, seeking peaches, confronts a fruiterer who refuses to sell them, leading to a series of encounters with a detective and a failed attempt to join a closed boxing club.
+  new scene 33 (LITTLE SPECK IN GARNERED FRUIT): The Kid seeks to expose gambling operations, challenging the captain to a risky raid on Rafferty's establishment.
+  new scene 34 (LITTLE SPECK IN GARNERED FRUIT): Denver Dick challenged the Kid personally, who outmaneuvered him in a game of wits and physical skill, then discovered a gambling setup gone awry.
+  new scene 35 (LITTLE SPECK IN GARNERED FRUIT): The Kid commands a nighttime raid on Denver Dick's restaurant, securing a peach for his bride, while the officers transport the prisoners.
+  new scene 36 (LITTLE SPECK IN GARNERED FRUIT): The bride desires her groom to perform a task, granting him power over her, while he fulfills her whims, including giving her a peach instead of an orange.
+  new scene 436 (WHAT YOU WANT): Old Tom Crowley, now a caliph with $42 million, seeks excitement beyond his wealth in a reimagined Baghdad-on-the-Subway.
+  new scene 764 (A HARLEM TRAGEDY): Mrs. Cassidy schemes to intervene in her neighbors' domestic conflict, while Mrs. Fink denies any physical assault by her husband.
+839 summaries: 829 carried from ../summaries.json, 10 new, 12s
+853 names in the corpus lexicon; 730 of 839 summaries changed
+wrote summaries_clean.json
+== embed summaries
+done: 839 vectors, 40s this session
+18:35:36
+== predicaments (qwen3:14b), all scenes
+839 scenes, resuming at 0, model qwen3:14b
+  10/839  30s
+  20/839  59s
+  30/839  85s
+  40/839  118s
+  50/839  152s
+  60/839  176s
+  70/839  205s
+  80/839  232s
+  90/839  257s
+  100/839  284s
+  110/839  310s
+  120/839  336s
+  130/839  362s
+  140/839  389s
+  150/839  415s
+  160/839  441s
+  170/839  467s
+  180/839  493s
+  190/839  521s
+  200/839  546s
+  210/839  570s
+  220/839  597s
+  230/839  623s
+  240/839  649s
+  250/839  675s
+  260/839  701s
+  270/839  727s
+  280/839  753s
+  290/839  780s
+  300/839  806s
+  310/839  834s
+  320/839  860s
+  330/839  886s
+  340/839  913s
+  350/839  941s
+  360/839  968s
+  370/839  995s
+  380/839  1020s
+  390/839  1048s
+  400/839  1075s
+  410/839  1102s
+  420/839  1127s
+  430/839  1153s
+  440/839  1180s
+  450/839  1207s
+  460/839  1233s
+  470/839  1259s
+  480/839  1285s
+  490/839  1311s
+  500/839  1337s
+  510/839  1362s
+  520/839  1387s
+  530/839  1414s
+  540/839  1440s
+  550/839  1467s
+  560/839  1492s
+  570/839  1518s
+  580/839  1545s
+  590/839  1570s
+  600/839  1596s
+  610/839  1622s
+  620/839  1648s
+  630/839  1674s
+  640/839  1701s
+  650/839  1727s
+  660/839  1751s
+  670/839  1776s
+  680/839  1800s
+  690/839  1826s
+  700/839  1852s
+  710/839  1878s
+  720/839  1905s
+  730/839  1931s
+  740/839  1958s
+  750/839  1984s
+  760/839  2010s
+  770/839  2037s
+  780/839  2063s
+  790/839  2087s
+  800/839  2113s
+  810/839  2140s
+  820/839  2167s
+  830/839  2191s
+  839/839  2215s
+done: 839 predicaments, 2215s this session; 219 still mention a job, relationship or place word
+853 names in the corpus lexicon; 187 of 839 summaries changed
+wrote predicaments_clean.json
+== embed predicaments
+done: 839 vectors, 39s this session
+19:13:11
+PIPELINE DONE