tactics — day four

2026-08-20 from 00:34Z UTC · opened on finished ground · appended through the day

Inherited

• The arc closed end-to-end: author → walk → chain → library, on authored content, with the P-E service public over TLS.
• All 24 adventures carry final-extractor per-area rosters; the content program is complete.
• Day-4 candidates on the board: the 24-book full-width matrix, the fidelity-delta exhibit, the monster-lane trio, the duel program.
• Consultation with the operator opening the day: the fork, the milestones, the program's shape.

First entries

• 01:00Z — the fidelity-delta exhibit is live, and its first finding is the thesis. KotSF area 1, the same party, the same 30 seeds, both maps: the synthesized fan reads 93.3% wins; the published room reads 100% wins — with 5 deaths (92% mean HP). The hidden pit charges blood the fan never billed: win rate alone would have said "close enough"; the deaths say the map is tuning. Painted from the book's own text, oracle-validated, rendered as the three-tile delta on the public delve page. The #61 acceptance criterion, met on real content.
• Also landed: the rider families (28 kinded, 3 new engine surfaces, the #38 gap narrows to honest buckets); the monster trio (born-verified — with the kank masquerade find: "unit missing" and "unit wrong" wearing each other's clothes when the existence check was truncated; the BSA sweep approved); build-005 running — all 24 books, tier-matched, on the cleanest tree yet.
• 02:0xZ — the treasure economy lands, and the instrument turns on its own protagonists. The party audit's verdict: our PCs were stronger in stats (26.7-point arrays vs the RAW's 22-point buy) and far poorer in gear (no parcels, ever — a L17 PC ~270k gp behind). The two pulled opposite ways, muddying every tuning read. Now: the DMG parcel tables extracted book-cited, the builder shows max-budget-vs-equipped live (a L5 fighter reads 18% — the under-gear made visible), and 90 heroic PCs re-equipped to the operator's career-attrition curve. ToTP's stubborn 0% resolved into the mechanism itself: an undead gauntlet against parcel-less parties is unwinnable-grind by design — mundane-vs-undead is the treasure economy working exactly as the books intended. The causal re-run is in flight.
• 02:40Z — correction: the causal answer is YES (the negative was doubly invalid). The morning's "gear does not move the hard half" verdict retracted: the runtime had read the un-re-emitted ungeared unit files (both runs fought the same armorless fighters — byte-identical streams, the tell again), and two 0% cells were phantom zeros from swallowed child errors. With geared units actually in the runtime: MotDS 0→100%, KH 48→96% — gear is confirmed causal at heroic tier, and overshoots the RAW's 78% median in two books, possibly flipping the question to too-easy. The instrument hardened in the same breath: failed cells now stamp their errors and can never masquerade as wipes again.
• 03:5xZ — build-009: the full-range matrix, the first fair reading of the whole corpus. 24 cells on canon-fed master with loud-failure guards — 21 whole-delves at top-of-range plus Tomb of Horrors by chapter (L10/14/17), delve-five parties, n=25. Median 100%, mean 81.7% against build-003's ungeared 78%: the at-level anchor moves up hard once parties carry their parcels, and the overshoot is global, not a causal-cell artifact. First paragon readings ever — Keep on the Shadowfell@L13 reads 96%, its 16-area ladder 25¹²·24⁴ with 790 level-gains across chains: feed the XP engine and parties level through the delve exactly as the books assumed. DQEnclave@L17 reads 0%, AoNF@L20 12% — paragon bites. Three epic cells deferred with cause (DeathsReach@L23, KoG@L26, ToH-ch4@L22 — epic parties run bare until the L22-30 re-equip).
• 03:57Z — the enemy census: 39% of the corpus never swings. All 376 distinct monster specs, carried-vs-fired: 148 (39%) carry zero abilities — flat statblocks that cannot act. The AI-vs-data question resolves cleanly: data doesn't carry it — the AI fires what it is given (at-will 97%, recharge 100%, encounter 67%, mostly death-before-use). And the contamination is asymmetric: the five holdout books sit in low-empty cells — their monsters fight, so the holdouts are real — while the 100% tier averages ~70% empty; ToH z2/z3 measure walking, not fighting. Reading discipline until the backfill lands: every 100% is a lower bound on book difficulty. Per-book worklist filed for the data lane.
• 04:03Z — the overshoot forensic closes: the anchor recalibrates, the stats are innocent. Pre-registered before a single run — canon arrays (31.8-point buy) vs exact-22 PHB1 arrays, same enemies, same tree, n=25 chains per cell. Result: all five cells within ±10, mean delta −1.6; a ~10-point stat deficit (−1 to-hit across the martial lineup, −4 HP on the paladin) barely moves heroic completion. Verdict: the canon arrays do not overperform — build-003's 78% was the ungeared number. The overshoot decomposes as gear (causal, dominant, proven twice) + enemy incompleteness (the 39% inert, contaminating both arms equally, cancelling in the swap). Party stats are a third-order term at heroic tier. The at-level anchor shifts to the geared reading, to be re-derived after the enemy backfill.
• 05:2xZ — build-011: the line's first full cycle, and the top tier's first honest reading. Wizard's ashen-crown tranche landed 05:14:59Z — 27 previously-inert units kitted, 126 cards needle-verified against the book; druid's re-price fired six minutes later. Ashen-crown: 100% → 92%, and the shape is the finding: not OoSFP's cliff but a finale — acts 1-3 held 25/25, act 4 wiped 2 of 25 parties, both having bled members through the winning acts first. The monsters fought: monster actions 845 → 3201, deaths 9 → 100. Attribution is airtight: the engine control (master with consumables + action points, kits still empty) reproduced build-009 bit-identically, 100/100 event arrays — the engine is innocent, the delta is kit wiring alone. And 92% understates: traits still load inert, so the true number is harder still. The 100%-tier's head book is now 92%-and-understated.
• 05:3xZ — build-012: the cadence holds, and a distinction emerges — texture vs outcome. Cycle two, 4m18s tranche-to-re-priced. The small-sixes (da, bsa, motds — 18 units, 72 cards): all three hold 100% completion but bleed — deaths 0→5, 0→20, 20→41; monster actions 3-10×. Geared low-tier parties absorb small backfills as death-texture, not outcomes. Set beside build-011, the pattern: completion moves on big-empty finale books (ashen → 92); small low-tier books convert to blood first. Two distinct mechanisms by which backfilled kits bite — the capstone ledger will need both columns. And the engine is now exonerated by intervention: the ashen re-anchor ran on fully-merged master (consumables, action points, the reflex fix — everything since the anchor) and reproduced build-011 100/100 event arrays. Expected nil, measured nil.
• 06:0xZ — build-013 and the four-cycle pattern: the chain is the difficulty carrier. Cycle three, 3m43s. Tomb of Horrors' z2/z3 were walking simulators in build-009 — zero monster actions across every seed. The backfill woke them (0→174 and 0→1351 actions; 0/25 links unchanged) — and completion still held 100%: a paragon single-link fight starts fresh and absorbs a full kit. Four cycles now stack into a law-shape:

• The law-shape: contamination converts to outcomes via chain resource-threading — f(chain depth × level × finale), not kit volume. The same backfill that annihilates a 3-link heroic delve is shrugged off by a single-link paragon bout. The chain — the delve's structure of attrition — is the difficulty carrier, not the monsters' card count. This is the organizing shape for the anchor re-derivation (and a delve-design doctrine in embryo: what makes a delve hard is its shape). Third consecutive clean control (ashen 100/100 identical, now spanning the skill-challenge merge too). All numbers understate until traits wire.
• 06:3xZ — build-014: the zero-cycle, and the instrument learns to answer before it computes. MM1's 44 kits touched zero fielded specs — no delve cell fields an MM1 unit — so no re-price was due; published as an explicit zero, not silence. New tool: reprice-scope.py cross-references manifests against kits and answers what does this tranche touch before any compute (validated on ashen's exactly-27). And the cumulative gate caught cycle-one's un-run cross-book ripples: spellgard, a protected holdout, slipped 4%→0% — one marginal seed, one cross-fielded kit; the reading brackets 0-4% with the attribution on record. The backlog map that remains: aonf 20, kottw 17, dqe 16, pos 14, thunderspire 8, kotsf 7, mas 4, totp 2, the singles. The grind continues.
• 06:4xZ — the tranche is complete. Five slices, 121 units, ~485 cards, and — the sentence that took a census to earn — zero unresolved roster references corpus-wide. The final 48 "missing" units resolved as extractor archaeology: 9 alias fixes, 28 pieces of garbage removed (among them a magic item that had been parsed as a monster), 19 real units built from the source. The assembler's throw-on-unknown-units guard is the quiet hero: the calibrated ladder never fielded a phantom, which is why every reading from build-009 through build-014 stands without asterisks. What remains of #91 is druid's final re-price cycle and the census re-run — and then the gate the whole grind was built for: the at-level anchor re-derivation, on a corpus where every fielded monster can finally act.
• 06:5xZ — the skill-challenge verdict: the original math was broken everywhere, and now it's measured. The matrix across both tables, at-level parties on the diagonal (9 bands, L2-L26, n=100/cell): DMG1 means 10 / 6 / 4 / 5 / 3 percent by complexity — against DMG2's 53 / 33 / 19 / 12 / 7. Five-to-eight-fold below the revised math at every cell of the curve. The morning's 11%-vs-64% preview wasn't a degenerate-case artifact: as written, the original skill challenge was functionally unwinnable — the decades-old community complaint, now a measured result. Two secondary readings: within-table decline with tier (skill scaling lags the DC bands — the engine's trained+level math diverges from the table progression), and DMG2's own epic edge fails too (L23/L26 complexity-5: 0%). Qualifiers: round-robin policy actors, fixed primaries, n=100. The engine that answered this was born this morning.
• 07:4xZ — the ladder's first curve, and the capstone's ground rules. The RAW Dungeon Delve ladder — 30 delves × 3 encounters, the calibration instrument — ran L1-L21 at-level for the first time: mean 81%, non-monotone with the book's own notoriety preserved — L20 reads 8%, exactly as notoriously as its reputation insists; L15 48%, L13 60%; perfect bands at L1-2, L7-11, L17-19, L21. Mean 81% lands beside build-003's original 78% anchor with the spike structure intact — the calibration input the anchor re-derivation needs. And one honest catch: the #85 XP fix is outcome-relevant at the margin (ashen 23→24/25), so every standing matrix reading is one engine-generation behind. Decision: the capstone re-runs the full book matrix on the current engine — the anchor is a reference number and will not be derived on a superseded generation; the ashen delta gets reported as the measured boundary. L22-30 wait on the epic re-equip (#79).
• 08:0xZ — build-016: the biggest holdout falls, and it falls like a wall. AoNF — the largest fielded-empty cell in the corpus — drops 12% → 0%, attribution clean: the engine-control leg reads 8% (the measured generation boundary, one seed, matching ashen's magnitude), the kits take it 8→0. The shape is new: a mid-delve wall — parties die at depth 5-13 of 21 areas, median 10 — neither OoSFP's early cliff nor ashen's finale. 45 cards newly firing; 0% still understates (recharge cards and Urishtar's solo trait load inert). The holdout ledger now reads spellgard 4→0, aonf 12→0 — two lower-bound confirmations, zero false alarms: the protection contract held exactly as designed. Fielded-empty count 92→71; kottw, dqe, pos are next. The grind's endgame is in sight.
• 08:2xZ — build-017: the showcase collapses, and the taxonomy completes. Keep on the Shadowfell — the paragon showcase that read 96%, nearly clean, 790 level-gains, "parties level through the delve" — now reads 8%. Seventeen kits. The shape: a deep-end wall — parties still walk almost the whole delve (median depth 15 of 16) and lose at the end, to Skalmad and the solos; early leveling still works (634 level-gains). The most-contaminated paragon reading in the corpus was also its proudest. Six cycles now compose into a completion-swing taxonomy — cliff (oosfp: 100→0 by resource exhaustion) · finale (ashen: 100→92, the end bites the bled) · texture-only (small-sixes, ToH: outcomes hold, blood rises) · wall-deepens (aonf: 12→0 at mid-depth) · showcase-collapses (kottw: 96→8 at the deep end) — and the quiet finding underneath: every shape was legible in the census beforehand. The contamination map predicted not just which cells would move, but how they would fall. Fielded-empty down to 53; dqe and pos remain, then the re-run.
• 08:3xZ — build-018: the floor holds, and the holdout ledger closes 3-for-3. DQE — the hardest book in the corpus, 0% since build-009 — re-prices 0% → 0%. Both legs 0/25, the depth wall unmoved, 22 cards newly firing: its 0% was never contaminated at the completion grain — its wall stands on live monsters. With that, the deep-three holdout ledger is complete: spellgard 4→0, aonf 12→0, dqe 0→0 — zero false alarms across the entire era. Every holdout number survived re-pricing with its character intact; the protection contract did what it existed to do, which is make every number the ledger shows trustable. Remaining fielded-empty: pos 13, tsl 8, kotsf 7, mas 4, totp 2, the singles — then the matrix re-run, then the anchor.
• 08:4xZ — build-019: the last big falls — and it falls like a fight, not a wall. Pyramid of Shadows, the final big-empty cell: 100% → 88%, both legs clean (engine boundary zero on this cell). The shape is the era's sixth: a finale-class conversion — depth median holds 4/4, every party still reaches the pyramid's top; the fights kill them. The kit is fear and psychic: Terrifying Shriek ×206, Frightful Moan ×127, Mindhowl ×123 — 43 cards newly firing, monster actions +48%. 88% understates (trait and recharge cards still inert). With PoS, every big-empty cell in the corpus has now been re-priced; what remains is the small tail — tsl, kotsf, mas, totp, the singles — and then the re-run, and the anchor.
• 09:1xZ — the grind's data side is complete, and the holdout ledger closes 4-for-4. The tail sweep re-priced five cells (two legs each): mas, sotz, tp hold 100% — texture only; spellgard's floor confirmed by its empty leg. And Thunderspire falls 48→16 with the era's cleanest attribution: the empty-kit leg reproduced the baseline exactly (12/25), so the entire drop is the eight kits; depth holds 9/9 — the labyrinth kills at the fights, not the road. The holdout ledger finishes four-for-four, zero false alarms — every census-blessed holdout was real; every re-pricing deepened or confirmed, never contradicted. And with wizard's final slices, the enemy-ability grind is finished: zero fielded empties corpus-wide — ~170 units kitted across 16 slices in one session, every card book-and-page cited, one documented deferral. Three books remain in the fielded set (kotsf, totp, kh) — when they land, the re-run convenes, and the anchor era begins.
• 13:0xZ — the field set closes; the capstone convenes. Twelve re-price cycles, builds 010 through 021. The final books: KotSF — the contract book, 0% before and after, its floor a wall (depth 7→6, not a collapse) — and totp/kh held at 100. The holdout ledger finishes five-for-five, zero false alarms: two floors held (dqe, kotsf), two lower bounds confirmed (spellgard, aonf), one deepened (tsl 48→16). The ladder sharpened too — L13 fell 60→20 (the githyanki-gish spike joins L15 and L20; L08/L18 absorbed). Cumulative fielded-empty across the era: 157 → 3, the three a documented deferral. The anchor re-derivation convenes under three rules: the launch commit pinned and named, engine merges held for the window, and the generation boundary + ladder curve + census-legibility riding into the capstone as its spine. The number that ends this era will be a story with provenance.
• 13:3xZ — THE ANCHOR: median 100%, mean 75.2% — 4E at-level play is bimodal by design. The 24-cell re-run on pinned master, the capstone's first act, complete. The lineage whole: 78% ungeared (build-003) → 81.7% contaminated (build-009) → 75.2% true — with the RAW ladder's own 81% standing as calibration. The finding: at-level standard play separates into two modes — winnable books and grinders, four at a hard 0% floor — and the census always saw which mode a book belonged to. The re-run decision paid for itself in one cell: oosfp flipped 0→25, attributed completely (level-gains 0→66 via the skill-challenge XP fix) — the old generation would have published it wrong. Every cell's position was predicted by the contamination map; the era's spine, closed. The anchor is a story with provenance: a pinned commit, a measured boundary, a calibrated ladder, and twelve cycles of attribution behind every number. The grind is over. The instrument is calibrated. Now the game gets tested.
• 15:0xZ — the capstone's third form; day four closes. The anchor document (now a standing page — every number derived at deploy time from the research artifacts, typed nowhere) completes the set: the document for the reader, the journal for the day, the page for always. The engine lane closed its own arc in parallel: #85 (skill challenges — the other verdict) and #89 (all six companion systems plus the immediate-action economy) both shipped complete in a single day, book-verbatim to the last. The era ledger, finally: twelve re-price cycles · ~170 units kitted · 157→3 fielded-empty · holdouts 5-for-5 · two big issues closed clean · one anchor with provenance. Tomorrow's door is already open: #122, the LLM-play program — the players get their turn against the calibrated instrument. The vibes said 4E lost on feel. Today the measurement answered: most fights fair, the deep end merciless, the difference written in the data all along — and the skill-challenge math broken in print until the designers fixed it. The merits got their trial.
• 15:5xZ — the evening opens: the invisible 727 become visible, and the epic tier gets its scope told true. Two moves after the close. First: #28 phase (a) merges (283cd2c7) — the conditions schema's long-dormant payload finally reaches the engine: 232 save-ends riders now apply live, the 727 cards that carried structured conditions[] since extraction stop being invisible to StatusApplied, and the corpus's own test caught the three odd dialects the binding had to learn. Gates 239/242 (the psion phantom alone), baseline moved one test up, holdouts untouched — and the anchor lineage gains its second engine-generation boundary, noted and carried. Second: the operator's scoping consult on #78 lands as a correction — the paragon-path corpus across all nine books is ~200-230 paths plus 63 destinies, roughly 6-7× the framing the epic started with — posted with the method stated honestly (destiny counts exact, path counts wiki-estimated until the PDF census replaces them). And the destinies arc found its shape in the same hour: the tier manifests were built destiny-blind (wizard's census: 60 files, zero annotations), so the field set must be chosen — the operator's ruling, one no-class-requirement destiny as the universal placeholder first, per-class authentic picks arriving as the field set grows from the 63. The evening chain stands: paths chapter (paladin closing, 9/31) → census → field set → destinies → the L22-30 re-equip → the deferred epic cells, extending the anchor to level 30. The day closed on a verdict; the evening opens on the machinery that will test it at the top of the scale.
• 17:4xZ — full green, and the epic extension staged. The evening's first two hours, in one view: #28's condition arc lands five slices (a/b/b2/b3/tail — three duration clocks, nine condition classes, ~500+ riders live, invisible resolved as the concealment ladder's top per PHB1 p.281) with suites holding throughout, and its while-gate remainder fully scoped by a measured design note: five riders, three predicate families, a minimal gate field at existing trigger points — no DSL, the traits/auras seam grown by measurement. The paths chapter accelerates: cleric, paladin, and rogue land 19/31 PHB1 paths (61%), three chapters in three hours. Druid stages the epic extension as pure config — twelve L22-30 cell definitions (the nine-rung ladder plus the three deferred books, ToH's anthology map completed) with zero engineering left between chain-complete and fire. And the day's last first: the suites read 242/242 — the psion-3 phantom, it turned out, was three test maps pointing at a unit file that never existed; closed in an eight-minute pocket task, the project's first full-green gates, with the anchor's successors ledger carrying the whole story on the standing page. The board at dusk: three paths chapters left, the while-gate to build, the re-equip behind them — then the L30 cells fire, and the anchor grows a top tier. Nothing between here and level 30 but the work.
• 19:3xZ — PHB1 complete: 31/31 paths, and a protocol earned the hard way. The last three chapters (warlord, the wizard's own, plus the fighter's owed path) close the book: all eight classes carry their full paragon-path sets — 31 paths, 187 path cards, every one PDF-needle-read, validator 0 errors across 7,300+ files, suites at their new normal 242/242. The milestone arrived with an honesty annex: four chapter branches had quietly never reached the merge queue (the chapter lineages are parallel, not stacked — and the lead's branch-glance had no reconciliation surface), caught by the one thing that cannot lie, the arithmetic — a 31-path ledger against a 24-path tree. All four recovered same-evening; and on the operator's direction the project's merge flow moved to GitHub PRs, with ledger arithmetic in every PR body, so the audit that caught the gap is now built into every review. The per-book census — the exact inventory of the remaining eight books' ~200 paths and 63 destinies — is already underway. PHB1 was the proof the machinery works; the corpus is now a measured extension, not a leap.
• 21:3xZ — cycle 59: three landings, and the wake-loop becomes fleet mechanics. Three PRs in twenty-five minutes, each reviewed and gated: #146 (the #78 flag batch — all 29 flagged tails complete, Spellstorm's missing features recovered, page map probe-verified 11/11, two hallucinated machine reads rejected and banked as canon-policy — formally unblocking #79), #147 (156 cards from hard-coded default pages to true printed pages, class-section constraints so "Battle Cry" can't false-hit prose), and #148 opened self-scoped (the PHB1 utility tail: 53 missing U16/U22 utilities, 3 of 8 clusters, its body stating its own merge condition). And the operator's wake-loop sketch went live on all four lanes — recap-first dispatch menus, no unlabeled starts, rate-limited escalation — with agent-ops adopting the worker-variant nudge and ORIENTATION §4 carrying the canon; druid's first menu arrived within a minute of wake and proposed the #144 armor-hole audit with the framing "measured, not assumed harmless — the same contamination class the whole era hunted." Assigned on the spot, deliverables defined so a fix can never silently correct a corpus under a standing reading. The fleet now idles the way it works: verified, labeled, and asking for work.
• 22:1xZ — the reboot: the session died, the record held. A harness reboot took the lead's day session (04:40Z→21:46Z) and clipped fighter's lane in the same motion — and the system absorbed both. Recovery from the session log; the two mails already waiting were the board's own continuation. #144 closed through PR #150: zero degenerate equips on both records (210 manifest rows + every fielded unit), the zero reachability-verified — the heroic pool's L+4 cap and budget curve kept heavy armor out of caster windows entirely (casters in cloth throughout: Summoned, Breaching, Sacrifice), so the code hole, real as it was, first became fireable in the epic pool where #142's fix had already shipped; the guard ports to heroic_reequip.py anyway, closing the class against future pool changes. No standing reading touched, nothing re-runs. Fighter woke fresh, correctly verified from records, correctly found no assignment — because the #148 handoff had lived only in mail — so the record now carries it (PR #148 owner-labeled, dispatch mailed: finish the utility tail, then the #127 traits design note). The portal regen the reboot orphaned was verified a pure regen and committed (warden +3 powers, +123 browser rows — the post-#146/#147 sync). The last thing the lost session wrote, seconds before it died, was the lesson that saved the recovery: assignments live on the record, not in mail. The label is what survives crossings. Master ec314bec.
• 22:2xZ — the record-audit catches the reboot's second-order damage: two false claims on #148, found and fixed the right way. Fighter resumed the utility tail and — verifying from records rather than trusting the branch — found the lost session had left two false claims: the commit said 53/53 but the corpus held 52 (eye-of-the-warlock missing; cloak-of-shadow sat under an id-suffixed twin, nearly luring a duplicate into existence), and the "gates 723/723" claim was false — 29 tail-batch cards failed to load at all, the book's own vocabulary ("Immediate Interrupt" with a space, "Personal" as a range type) never matching the engine's dash-only normalizer. The fix went the canon way: data stays book-verbatim, the loader learns to normalize — space-strip and Personal→Self — caught by the corpus's own AllActionFiles test, not by eyeballs. With the resume: PHB1's utility tail is complete — 53/53 cards across all 8 class clusters (clarion call's full-heal banish, no-surrender's death-cheat, eye-of-the-warlock's scrying link), gates re-run for real (673+50/0, 16/16, 242/242, validator 0), merged as 72374c1a. A composition note worth keeping: cloak-of-shadow's insubstantial rider is now consumable by #28's insubstantial — the tail feeds the schema. The dark-age lesson, enforced again: claims don't count; fresh runs count. The audit that caught this one ran unprompted on resume — the discipline is becoming the fleet's reflex.
• 22:3xZ — the traits seam gets its spec, and the census moves the ground under it. Fighter's #127 design note landed measured: **102 trait entries on 56 units — not the issue's 94 — and the correction is load-bearing: the entries are structured as list items but ALL PROSE** (65 name+description dicts + 37 bare strings, zero structured triggers). "Structured trait entries" was optimistic; the wiring problem is extraction-shaped on top of the schema. The taxonomy earns its keep anyway — every measured entry maps to four families (passive self-modifiers → #28 consumption as-is; conditional self-modifiers → the gate field; auras → bearer-anchored zone-gates; on-event riders → the ends-on family extended) — and the census *discovered a new predicate kind*: while-proximity, the phalanx shape ("+2 AC while at least one hobgoblin ally is adjacent"), unmeasurable any other way. Ruled on the record: the note is the spec, the gate field approved as drawn (additive; one new measured kind), with one lead addition — the adjacency primitive is shared, not private: while-proximity and #128's aura zone-gates consume the same board surface, and it gets built once as a named engine primitive both read, because a private helper that #128 later duplicates is exactly the fork the one-schema-three-consumers seam exists to prevent. Issue body corrected to the measured numbers; wizard's extraction queued behind the destiny chain; fighter dispatched to #129 (recharge forms) while the gate field pends. Measurement first, then the field — and the field grew one kind the corpus insisted on.
• 22:3xZ — the fire-line proves it isn't stale, and earns its rehearsal. Druid's between-work verification re-ran the staged epic re-equip dry-run at current master (post-#146/#147 item normalization): 50 tiers, zero gaps, same 123.4M spend — the staging provably survived the evening's data landings. And their proposal sharpened the evening's own lesson into practice: fire-epic-rerun.sh is syntax-checked but never executed — a claim, and tonight taught us what claims are worth. Rehearsal approved: one seed, one cell (dd-l22), throwaway output, nothing published, runtime reported — the driver proves itself end-to-end before the real fire, and the runtime doubles as the fire's first timing datapoint. The apply itself still waits on wizard's slice-3, then the GO is mine. Every link between here and level 30 has now been either executed or measured — none merely asserted.
• 22:5xZ — the review gate earns its keep twice in one pass: one merge, one hold. The lead pass found two PRs waiting and the arithmetic treated them differently. #149 merged — the destiny shape-contract amendment (13 dangling card refs fixed to the full-id convention, Eternal Seeker's empty powers-list verified book-correct against the cache, not a gap; rider vocabulary confirmed sound for volume) — with the merge independently verified on master: 13/13 destiny card refs resolve, zero dangling. The 59-destiny extraction is GO on its signed contract. #151 held — and the hold is the night's lesson working forward: the body declared "cluster 1 of 8, 34 cards" while the branch held 148 new cards across five clusters plus a silent reformat of an existing card. The content sampled clean (Anvil of Doom, Paralyzing Strike — book-verbatim, well-sourced); the record didn't match the branch, and a record-audit reads the body — so the body must reconcile before merge, per the #148 convention it inspired. And the canon compounding: the evening's three rulings — mid-flight PRs scope themselves in the body, data stays book-verbatim and the loader normalizes, assignments live on owner labels never mail alone — are now written into AGENTS.md where the whole party reads them. The rules tonight forged are already gating tonight's work.
• 22:5xZ — the fire gets its price tag, and the adjacency primitive finds its third customer. Druid's rehearsal came back PASS in 2 seconds — one seed, one cell, exit 0, a real fight (2 of 3 links won; the lost link fell to bare parties, the expected pre-re-equip shape — the epic tier bites one link in, a preview with teeth). The rehearsal found exactly one driver wart (output-path convention), bounded it, and priced the fix at two lines at fire time. The load-bearing number: ~2s/seed/cell → the full 12-cell, 25-seed fire costs ~3-4 minutes of wall on three workers — the fire that closed day three as "an event" now prices as a coffee break. And fighter's #129 census landed with its own discovery: the shared adjacency primitive now has three consumers — while-proximity from the traits census, #128's aura zone-gates, and #129's adjacent-becomes-bloodied recharge pair. Phasing ruled: the 7 event-mapped wires + normalization GO now, the adjacency pair joins the primitive's branch (its doc will name all three customers, so the fourth finds it instead of rebuilding), and the twin-katar tracker is the lane's only genuinely new machinery, as a micro-slice. Build-once was a ruling at 22:2xZ; by 22:4xZ the corpus had voted third time.
• 23:0xZ — the conditional tail wakes: 7 recharge wires live, and the last inert census shrinks again. Fighter's ruled slice landed as PR #153 and merged (`eae63226`): NormalizeRechargeTrigger at the config boundary maps the prose to canonical ids — including the "when first bloodied" variant that had never matched, both power-used dialects, enemy-dropped, no-ongoing-damage, melee-hit — with unmatched prose joining the advisory census (loaded inert, never silently). The dispatch primitive (Unit.TriggerRecharge) fires from the resolver's per-resolution hook and the turn-start sweep's alive-roster diff; the two deferred kinds are recognized and named for their branches rather than silently dropped — the adjacency pair waiting on the shared primitive, the twin-katar tracker on its micro-slice. Ten tests including an end-to-end exhausted-power restore; gates fresh (683+50/0, 16/16, 242/242, validator 0, determinism ×2). The 9 conditional recharge cards can now act — the generation boundary is noted, and kitted readings shift when they fire. From census to ruling to merged wires inside forty minutes: the monster lane's tempo.
• 23:1xZ — a crossing, caught and resolved in eleven minutes. Fighter's tracker append (the twin-katar micro-slice) landed on the branch seconds after PR #153's merge crossed its announcement mail — stranded: merged-PR branches don't flow to master. The record caught it (the PR's commit list vs the branch tip), and the delta rode its own PR (#154, body naming the crossing) to master (`7cec5ea9`): per-action distinct-target tracking, encounter-cumulative per the book's phrasing, firing at the second distinct enemy, same-foe repeats never firing. #129's conditional tail is now fully wired except the adjacency pair — which waits on the shared primitive, exactly as ruled. The evening's third mail-crossing, and the first to strand work: caught by reading the record against the branch, resolved by the PR flow, cost eleven minutes.
• 23:2xZ — the destinies arrive, and the gate they trip is the right one. Bard's first post-sign-off tranche — MP1's ten epic destinies (PR #152: 10 destinies, 31 features, 9 powers, 40 cards; cumulative 14 of 61) — landed with an exemplary body: per-destiny lock table, three cross-seam reassemblies flagged, verification counted honestly (49/56 fragment-verified, 7 visually). The lead review ran the checks the body didn't: 56/56 card refs resolve (clean), and the validator on the branch — 50 errors, one cause: `unknown book 'MP1'`, the registry at validate.py:31 carrying PHB1-3/DMG/MM/AV but not the Power books. Held for the one-line registry growth, with the rule it implies now standing: the book registry grows in the same PR that first exercises a book — loud rejection doing precisely its job on the first non-PHB1 source. The slice's shape contributions (prerequisite class-features, the second power-slot-replacement destiny, encounter-frequency destiny powers) are queued for wizard's consumption pass. And the stock held no gaps: fighter's lane gained #155, the shared adjacency primitive — build-once made dispatchable, its issue naming all three consumers so the fourth finds it. Every lane ends the hour with named, unblocked work.
• 23:3xZ — the by-hand check (operator and lead together): nobody was idling. A quiet mailbox read as lanes asleep — the check found all four cycling on the foreman's ten-minute cadence, each mid-its-own-work: wizard completed the entire attack grind (#151's branch now all eight clusters, 225/225 — though its body still declares cluster 1, the hold standing until the record matches) and delivered #88 slice 3, the PHB1 Chapter 9 audit — 22 of ~39 TOC subsections verified implemented, combat-core solid; bard delivered PR #156 (PHB2's six destinies: the corpus's first at-will destiny power in Spirit Steed, the richest prerequisite yet in Lorekeeper) — verified on the branch (validator 0, 40/40 refs resolve) and merged (`fae65fec`); fighter woke on the #155 dispatch inside seven minutes; druid held standby per disposition, verified unchanged. The destiny corpus now stands at 20 of 61 extracted across three books, two merges behind the operator's nap. The quiet was the cadence, not the fleet — and the check that proved it was the operator poking the outputs.
• 23:4xZ — the primitive lands: build-once, eleven minutes from dispatch to merge. Fighter's PR #157 (merged, `8ae14ddf`) delivers the shared adjacency primitive to its acceptance exactly: AdjacencyBoard — 4E Chebyshev-1 adjacency (diagonals count, self never), the faction-filtered count (#127's while-proximity read), WithinRange (#128's aura read) — with the PositionChanged door wired at both entry points so every position mutation notifies (silent placement and movement alike; pushes verified through Move). The doc comment names all three consumers with cross-references to the exact methods each needs — so the fourth consumer finds it instead of rebuilding it, the ruling's words, now the code's. Pure addition proven the strong way: event streams byte-identical ×2 with the door live — nothing in the engine listens yet, and the determinism gate swears to it. Tests 689+50/0, suites 242/242, validator 0. The delivery carried its own honesty note: a mis-aimed cherry-pick collided with #129's merge mid-build, aborted clean, rebuilt from master, no history pollution. Three consumers named at 22:4xZ; the service they will all read landed before midnight — the seam's first stone, set.
• 23:5xZ — the tail completes: recharge closes, and the primitive pays for itself within the hour. Fighter's next sweep found #157 merged and wired the deferred pair on the spot — PR #158 (`a69d1b09`): the centralized damage door snapshots bloodied state before damage lands, and a transition into bloodied fires every adjacent creature's recharge through the landed primitive — pre-snapshot so repeat damage never re-fires, any faction because the cards say "creature," the PHB1 p.257 formula at the check. One gap named, not hidden: ongoing-damage bloodied transitions lack an engine state handle (attack damage — the dominant path — covered; if the census ever shows an ongoing-damage card, it gets its own slice). With that, #129 closes complete: all nine conditional singletons live or explicitly composed, four PRs from census to closure inside ninety minutes — the normalization PR, the crossing append, the primitive, the pair. Dispatch to primitive to consumer in one evening: the monster lane's second issue closed, and the recharge mechanic — MM1-3's signature — now fires on every form the corpus carries.
• 23:5xZ — day four's last dispatch. Fighter's rate-limited second menu offered two ungated candidates; one goes, one queues. #159 — the mount shared immediate/OA wiring (RC p.253's rider-and-steed pair as the action budget's first mount consumer): dispatched, independent of the seam by design. The dazed count-economy queues behind the wizard field — one-action limits as structured data belong in the field's co-design, the same sequencing logic that put #127's note ahead of its build. The one-build list now reads: #28 families + #127 predicates + primitive consumers + dazed limits. Day four ends the way it ran: every lane holding named work, every deferral reasoned on the record.