Next 16.3.1 turns on turbopackFileSystemCacheForBuild by default and keeps it in
.next/cache/turbopack, inside the image build's BuildKit cache mount. Reproduced
locally: a cold build at e8e358213, then a build at 0bf3210c8 reusing only
.next/cache, lost the four .birth-place-coordinate* classes exactly as CI did.
With the cache off the same rebuild carries every class; Turbopack compile time
is 21.0 s either way. Recurrence of BUG-2069, whose cache-id rename only reset
the cache once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The astrologer's 11-21 reply (owner agreed): single positions cannot be turned
into personality, money or family conclusions, nor can resemblance picks narrow
a birth time. describe() now emits facts only; resemblance questions are retired.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2U3BH1SxwAGyhhbe1
Astrologer's Yoga manual 2026-11-21, P0. asraya_partial and asraya_yoga are
disabled with an accuracy note (ids kept); Aasraya proper stays as
Rajju/Musala/Nala. Two Kaalanirdesat progeny rules get conditional Chinese and
English effects. Topics file rebuilt; enabled count 408 -> 406.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2U3BH1SxwAGyhhbe1
Owner 2026-11-21: after BUG-2115 the full-data report still printed source
lines and book references (Goel 出处 line, the PVR 2014 article in the
conditional-dasha note, PVR / Light on Life / BPHS verse in the calculation
notes, Tithi lord table source, rashi-dasha 出处 lines, Narayana p.43 label and
p.50 / Table 13 direction notes, functional profile source row, Ashtakavarga
book calibration). Method names stay; the book, page, chapter and ruling round
go. Engine labels for the consultation card are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2U3BH1SxwAGyhhbe1
A host-side systemd timer (04:45 Asia/Shanghai) makes sure a restore-verified
daily backup exists (running the existing recovery script in daily mode when the
Gitea workflow did not), then uploads it with an OSS V1 signature computed in
the script, Content-MD5 and an ETag check. Secrets never reach argv or output.
Failures write a marker and alert through the existing Feishu monitor.
Not installed on the production host.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The reading facts gave the writer 「PVR 表 30」 as the functional-profile label and
Narayana readings with page and table citations; the annual material and its
program-written notes named author versions and table numbers. Facts now carry
plain labels (ids kept), every report writer and in-report answer gets one
no-citation clause, the reading and the answers are checked, and stored text
loses its citations where it is shown.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The reading facts gain timing.period_connections from jyotisha-20's indirect_activation (BUG-2121): per
house only the matched paths, their lords and levels, and one fact line; built even when the data
chapters were not rendered. The writer rule and the question instruction call it an index, not an
activation verdict; the eight body sections hold back 激活 / 引动 / 被点亮 / 被触发.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
A security-definer function, gated by support.quality.read and granted to
admin_runtime only, returns per report the reading's section statuses and
the transform log without any reading text; the admin page 报告写作记录
lists sections, reasons, repair, salvage and fallback per report.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Annual clock captions no longer end with "this report's native calculation";
the general dasha reference drops its per-lord study-notes citation; the birth
data section translates user_reported / family_exact and the other stored codes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2U3BH1SxwAGyhhbe1
Round twenty-four: 「事业」 and 「感情」 fell back to three bare computed lines. A section
was sent back with its first reason only; the repair fixed that one and the dated
sentence without its birth-time condition survived. The condition is now inserted by
the program, the repair note names every reader-wording rule, and after the repair a
section failing only on reader wording drops the offending sentences and keeps the
rest when at least 60% and two sentences stay. Also: richer computed 事业/感情, the
basis rule names the Narayana, witness and same-fact fields, report scenario note and
annual Bhrigu note in the 11-21 wording, and a claim about every planet's strength is
unsupported when the material lists one below the line.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The annual method data shows only its numbers and clock facts; the research
chapter says "reference algorithm"; KP and Sade Sati describe the ephemeris
without naming Swiss Ephemeris. Reader leak checks now cover the whole report.
BUG-2055/2058/2063/2064/2074/2079/2090/2102 marked resolved after the owner's
staging walkthrough.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXp2U3BH1SxwAGyhhbe1
Per the astrologer's 2026-11-21 review (items 1 and 2): one line per
section in her wording, naming only the affected claim types; a dated
conclusion in an affected paragraph says it is by the chosen birth time;
question material keeps the line; stored readings show the new line;
dignity is never mapped to 施展不开 / 最能发挥 / 问题已经消失.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The astrologer (2026-11-21) ruled that ordinary readings and forecasts never
use the user's experiences or feedback; only the rectification flow may.
Her reference files still told the model to back-calculate the user's known
events first, stop predicting below 3/5 matches, and grade A by such
verification. Until she edits them upstream, a product layer cuts those
passages out of every text sent to a model: the personal reading's quoted
references, the chat's methodology excerpts, and every chat prompt before
it reaches the provider (covering skill_read). The files are not edited;
a test names any rule that stops matching once she changes them.
Grade A no longer asks for a historical back-test. The status
user_history_verification_required is no longer a gap: no verification
questions, no penalty in ranking, never shown as 「需要过往经历核对」, not
offered to the report or synastry writers. 「未见阻断项」 is caught in the
reading and in report-question answers. Contract §AU announces the layer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
Rules sharing an English name printed each other's Chinese name in the reading
facts; each row now takes its engine row's name_cn by rule id. Neecha Bhanga rows
carry the fallen planet and the cancellation witnesses as two roles in the data
chapters' words. Narayana's direction readings are read from the report packet
when BUG-2105 stores them, and skipped otherwise.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
After BUG-2100 merged the supplementary timing chapters, one chapter holds up to
~505k characters and a question read only its first 20k. A chapter over 12k
characters is now read as its lead, a list of its ### parts by id
(chapter_<n>_<k>), then parts from the top; a part can be read by its id, and a
passage selected in a long chapter is read from the part that holds it. Cut
material ends with a mark the model is told about. The TOC timing pattern gains
年运 and Timing so 「年运附表」 and "Additional timing calculations" match their
other-language twins.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8
The personal report reading wrote UL, 入庙, 小运, 婚姻煞 and the like in the
sections written for ordinary readers. Owner 2026-11-20: the body uses plain
words, technique names stay in 「依据与边界」 and the data chapters.
- Rules (product contract, report-reading block, notice G-9): the eight
sections take the chat's PLAIN_TERMS and time-naming rule; dignity is judged
from dignity_zh but told by its meaning; 「现在这几年」 gives dates and the
ruling planet, no dasha names, no Narayana; names only in 「依据与边界」.
- Check: technique_name, after every fact check, through the existing one
section repair; a writer's own plain gloss next to a name is kept instead.
A section held back for a name still counts toward the draft length, so the
draft stays usable and reaches the repair.
- Computed fallback text uses plain labels and drops the Narayana paragraph.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DbBXoyn7Hj3LCW3BUs5bZ8