lotraAI Image2Slide 1.37.5

lotraAI Image2Slide 1.37.5

1.37.5 基于 1.37.4，修复原图重建中的执行恢复与证据身份问题，并把字号、字体家族和居中检查接入普通组件工作流。文字、表格和公式显式字号使用统一 pt，按 0.5 pt 十进制就近舍入、中点向上；源字形与真实绑定字体进行有限物理字号拟合，原生写入前检查、写入后立即读回属性。中文与英文使用各自有限字体集合，先识别视觉家族，再按有界强证据细分实际字面。整段墨迹位置在横纵两个轴检查，保留原图有意偏移。

原图语义内容能够可靠界定时，可选择内容轮廓；否则保留全图恢复。原始字节及可逆映射始终保留，语义裁取不自动建立透视校正或原画布比例。用户委托按图判断时，可保留多处文字/边界推断、假设、不确定性和留出复核后选择估计工作平面，不能冒充原稿元数据。字号预检失败会返回合法短语采样恢复包，保留暂存字体、预算和失败记录；无支持或有歧义的拟合不能冒充唯一正确字号。事务恢复、惰性状态协议、源线支持、可编辑材质、逻辑/原生文字和首次渲染局部闭合共同进入最终检查。

当前回归覆盖：Windows 523 项通过；Linux 522 项通过、1 项 Windows 专属跳过。历史 R3 私有夹具错误及修复记录保留。当前 R4 双平台完整回归重新执行，Linux 本轮首次安装因自有测试盘空间不足失败，归档自建临时副本并预置同哈希资源后从正式入口恢复，初次结果保留。Windows/Linux 最终载荷首次和重复安装通过，重复安装 475 个技能文件字节和时间戳不变。实际分发入口通过 70 次解析，真实完整包、嵌套载荷及双仓库版本/内容检查 2564 项通过。

原事故由官方 DeepSeek Harness、现有账户和 deepseek-flash/max 实际生成原生候选并用 PowerPoint 渲染，完成了多项局部修正。本轮箭头诊断已闭合：见证点是最坏像素差而非箭头中心，改用较小原生箭头后，走线、端点、拓扑和实际涂料检查通过，源路径和限值保持。本次发布范围为当前诊断闭合后的产品版本；原事故当前仍为 review 检查点，没有整图验收、独立最终审计、正式导出或 delivery-check allow_finish:true，其余组织框/线路/分组外观、布局、延后字体与整图遗漏检查保留为未完成。局部诊断和原生/单位检查不构成整图 PASS，不推出总体 token、时间或多图质量增益。原图、候选 PPT 和私有字体/任务数据不随产品发布。

产品版本 1.37.5 与 workflow 1.35.2、design revision 8、Math 策略 1.35.3、runtime resources 1.14.0 独立。未变的资源和 18 个历史分块复用，只有 1 个新尾块；没有新增运行库二进制 Release、全局技能或模型密钥。主项目自定义 1.37.4 注册保留。

Windows 当前使用 PowerPoint，Linux 当前使用 LibreOffice；WPS/macOS 没有对应实机资格。历史 Office 字体再保存失败、原始 LibreOffice Math 分支/样式/几何保留限制继续单列。公开 README 命令和完整新尾块下载的首次/重复结果，只在正式发布后实际执行并记录到 Installation-checks.json；本发布前报告不把尚未运行的公开步骤计为通过。

lotraAI Image2Slide 1.37.4

lotraAI Image2Slide 1.37.4

Based on published 1.37.3, this release repairs the font preparation contract and
its interaction with source text, native slots and finite component decisions.
The product remains a self-contained project skill.

Every ordinary run and table cell gets a complete verified latin/ea/cs plan.
Writer and expected-style checks share its resolution; independent saved-byte
checks still verify original literals, logical colour/emphasis intervals, actual
slots and bound resource identity. A Chinese binding that differs from a font
hint no longer produces false unused-slot failures. Mixed-script and genuinely
different source families remain representable without merging coloured runs.

New tasks use font-workflow-1.37.4. Ordinary font hints no longer define binding
identity. Exact task/source requirements, source-intent revisions, true weight/
slant requirements and user locks remain authoritative. Candidate filtering and
binding share feasibility rules. Synthetic style requires its own evidenced
decision. A source-supported style group adopts one family while each span binds
its appropriate real face. No global font preset or page-majority assumption is
introduced; inferred photographic identity stays approximate.

Font packets visibly name their finite preparation step, batch and request IDs.
Source style, resource decisions, isolated glyph samples and family choice
precede binding. Related submissions apply atomically and return explicit
receipts. Exact acknowledged retries read the receipt. Native adoption remains
owned by the existing saved-file journal. Explicit old-task migration retains
old identities/history and cannot downgrade exact font requirements to hints.
Runtime comparison registrations survive detached pending ledgers and actual
submit/retry. Group revisions may span bounded packets; old faces stay unapplied
and block native adoption until every current group member is ready.

Source-line failures return legal source/native decision work. Unknown or
incompatible component states stop with a diagnostic instead of auto-evaluating.
Before construction, the actual writer's unsaved native emission is independently
decoded and checked against frozen line defaults, routes, topology and terminal
facts. This preflight does not replace saved-byte, render or final line gates.
An original-only forward run also exposed a missing correction exit between
source sealing and component binding. The binding packet now accepts bounded,
prior-hash source corrections with current observation/checkpoint identities;
the full original validators, obligations, bounds and observation history remain.
After sealing, a focused source decision can also correct its existing source
feature rows with exact prior hashes and then requires an explicit rebind. Real
open-shaft contacts use line_junction with original incidence and measured peer
geometry, without a hidden shape port or automatic-rerouting claim. Rejected
preflight reports remain registered hash-bound task dependencies.
Current component packets expose the current candidate's exact saved references
even before a selected baseline exists. Local compilation retains untouched
legacy nodes only when effective slots and bound resource identities agree;
real outside-scope font, resource or emphasis changes still require a grant.
Font discovery descriptors retain immutable project-local catalogue snapshots.
A render command with another backend configuration cannot replace a task's
font environment. Legacy shared-catalogue replacement triggers the normal
capability refresh; changed resource bytes and snapshot tampering stay blocked.
Font sizes quantize once to native hundredths of a point for ordinary text,
mixed-Math paragraph defaults and tables. This prevents the observed 18.15 pt
source intent from serializing as 18.14 pt through an intermediate EMU floor;
saved-size checks continue to read the native integer independently.

Line raster refinement retains the same fixed occupied source/render cells.
Exactly axis-aligned paired width chords can use rational cell intersections
instead of an irreducible half-source-pixel sampling floor. Every chord site
must correspond; missing sites, clipped support and unsupported directions
remain explicit. Source-to-render shaft coverage can refine its existing owned
cell set with finite interior sampling. Coarse inconclusive probes stay in the
report; valid failures and two-direction terminal/contact checks remain binding.

An independently saved, opaque polygon marker with no outline/effects now owns
its measured paint envelope. Its decoded vertices, actual fill and source anchor
bound the comparison plus the unchanged source tolerance and contact guard.
It no longer inherits a shaft-width multiplier intended for native line endings,
which had pulled unrelated text into marker checks. Other unqualified paint keeps
the conservative producer. Oversized/translated markers enlarge the measured
extent and still fail; internal fill and neighboring contact paint stay checked.
Separate marker anchors use their own saved vertices rather than substituting
the shaft endpoint. Previous full-crop differences remain retained under the old
producer; they are not relabeled as current marker PASS.

The 1.37.3 Windows verified font-stage resume, sharing-lock recovery, prior binding
hashes, scoped source grants, user locks and exact managed dependency reads remain
protected by regressions. Source workflow 1.35.2, design revision 8, execution
policy workflow-execution-1.37.0, verification policy evidence-checks-1.37.2,
native Math policy 1.35.3 and runtime resources 1.14.0 remain independent.

Both repositories use dev and 1.37.4; the formal source tag is v1.37.4. Unchanged
runtime assets and immutable complete-bundle parts are reused. Actual measured
results appear in the current validation receipts. Synthetic fixtures do not
accept external customer artwork; physical WPS/macOS require their own host
evidence. No token, speed or reconstruction-quality gain is claimed.

The customized active installation, user sources, private tasks/fonts/knowledge,
existing local edits, prior tags/assets and global skills are preserved. The
clean release source is an isolated maintenance checkout.

Current measured release evidence

Windows: 350 regressions executed and passed, including 85 font
cases and the actual external open-ZIP sharing-lock/release/resume test. Linux/WSL:
349 executed and passed out of 350 discovered; the sole skip is the
Windows-only sharing test. Final metadata and archive checks
passed 2477 consistency checks and 70 actual final-file parser checks.
All five bounded-dependency read cases passed on each host, including exact
returned children, pagination, cache replay and rejection of changed or escaping
references. Final installed manifest and read-layer hashes bind the task evidence.

Both current final installations passed first/repeat, 1.37.3 to 1.37.4 differential
update and repeat. The update preserves 417 byte-identical managed files
and their mtimes and changes/adds 41; all 458 files remain unchanged on repeat. Managed local
edits and untracked user data block an update and are retained. Runtime resources
and immutable archive identities are unchanged. The bundle reuses 18 immutable
parts and supplies one new tail. Public exact README and full download first/repeat
receipts are completed after tag and channel activation.

Fresh installed task fixtures passed saved rendering, current visual review,
independent final gates, formal two-file export and exact-byte delivery checks on
Windows PowerPoint and native Linux LibreOffice. Fourteen delivery counterexamples
per host were rejected. Nonzero-line replay and a separate original-only forward
test use this implementation, with new current observations and no historical PASS.
The 24-case renderer calibration on each host retains raw preset divergences;
production checks reject those divergences and all 24 explicit editable native
routes meet the unchanged limits. Raw-preset calibration is not relabeled PASS.

The independent forward run found and retained pre-bind source correction,
open-shaft junction and diagnostic-read blockers. Current product fixes and their
counterexamples were applied through the maintained installer, and the same
original task resumed. The superseded unpublished package and its earlier
receipts remain separately retained; none is relabeled as final-payload proof.
These fixtures do not qualify arbitrary Windows path depths, physical WPS/macOS,
external customer artwork, or token/speed/quality gains.

lotraAI Image2Slide 1.37.3

lotraAI Image2Slide 1.37.3

This maintenance release fixes font revisions that could remain blocked after a
valid component source repair, Windows font-package commit recovery, and saved
run/cell font verification. The product remains a self-contained project skill.

An existing font revision packet now names its exact owner, location and script,
the prior binding hash, source font field and current authority or blocker.
Whole-owner authority covers that owner's font revision; a physical text-body
write alone does not grant unrelated font changes. Existing user locks and exact
font requests remain authoritative. A located evidenced redecision retains the
unapplied plan and returns affected owners to source work. Binding decisions stay
detached until the established native journal saves the actual file.

Direct EOT and provider font stages retain their owner, input, binding and staged
byte identities. Source, output and verification ZIP handles close before replace.
A Windows sharing lock retains the verified stage and unchanged input. Recovery
continues that commit and its saved build instead of reconstructing geometry or
spending another native-construction attempt. Changed or unregistered files are
retained as conflicts. Commit retries have a finite budget. Completed replacement
can resume after interruption before the next caller journal transition.

Saved binding checks resolve owner/location/script to each actual run or populated
table cell and verify its literal and font slot, including mixed Office Math.
Swapping two families within one owner no longer satisfies the binding gate.

Product 1.37.3 retains source workflow 1.35.2, design revision 8, execution policy
workflow-execution-1.37.0, verification policy evidence-checks-1.37.2, native Math
policy 1.35.3 and runtime resources 1.14.0. No new runtime binary release is needed.
Both repositories use dev and 1.37.3; the formal source tag is v1.37.3. Existing
immutable runtime assets and unchanged bundle parts are reused.

Validation includes current Windows/Linux regressions, real Windows file sharing,
current installed-payload task/delivery fixtures, installation first/repeat and
download/README gates, final ZIP/nested-payload version and parser checks. Exact
counts and results are recorded in the accompanying current validation receipts.
Synthetic fixtures do not accept an external original. Physical WPS/macOS need
their own host evidence. No token, speed or reconstruction quality gain is claimed.

The existing customized active 1.35.3 installation, original images/PPTs, private
tasks/fonts/knowledge, local edits, prior tags/assets and global skills are retained.
Clean published source is the isolated development-v1.37.3/source checkout.

Current bounded managed reads now deliver the exact child dependencies returned
by an already authorized frozen parent. This closes source Scene and line
measurement reads needed by normal repair/font workflows. Only the selected
page's verified same-task/source resources gain a lease; unselected siblings,
stale/changed bytes, another original and task-escaping paths remain rejected.
Five focused regressions cover this delivery boundary and its cache behavior.
Constructor/acceptance rules, implementation-bound native results and quality
thresholds are unchanged by this read-delivery repair.

Current measured release evidence

Windows: 245 regressions executed and passed, including all 32 font cases and the
actual external open-ZIP sharing-lock/release/resume test. Linux/WSL: 244 executed
and passed out of 245 discovered; the sole skip is the Windows-only sharing test.
The Linux font suite executes 31 portable cases. Final metadata and archive checks
passed 2445 consistency checks and 70 actual final-file parser checks.
All five bounded-dependency read cases passed on each host, including exact
returned children, pagination, cache replay and rejection of changed or escaping
references. Final installed manifest and read-layer hashes bind the task evidence.

Both current final installations passed first/repeat, 1.37.2 to 1.37.3 differential
update and repeat. The update preserves 435 byte-identical managed files and their
mtimes and changes 18; all 453 files remain unchanged on repeat. Managed local
edits and untracked user data block an update and are retained. Runtime resources
and immutable archive identities are unchanged. The bundle reuses 18 immutable
parts and supplies one new tail. Public exact README and full download first/repeat
receipts are completed after tag and channel activation.

Fresh installed task fixtures passed saved rendering, current visual review,
independent final gates, formal two-file export and exact-byte delivery checks on
Windows PowerPoint and native Linux LibreOffice. Fourteen delivery counterexamples
per host were rejected. Nonzero-line replay and a separate original-only forward
test use this implementation, with new current observations and no historical PASS.
The 24-case renderer calibration on each host retains raw preset divergences;
production checks reject those divergences and all 24 explicit editable native
routes meet the unchanged limits. Raw-preset calibration is not relabeled PASS.

An initial nested Windows test layout exceeded native dependency DLL path limits;
its failed task and logs are retained. A fresh short owned project ran the same
final payload. A manually reconstructed forward-test invocation omitted the
registered UTF-8 flags; its failure is retained, and accepted calls use the exact
Bundle-install command_prefix. Neither failed attempt counts as product PASS.
These fixtures do not qualify arbitrary Windows path depths, physical WPS/macOS,
external customer artwork, or token/speed/quality gains.

lotraAI Image2Slide 1.37.2

lotraAI Image2Slide 1.37.2

The baseline is published 1.37.1. New tasks bind evidence-checks-1.37.2:
independent source line contracts, program-compiled obligations, saved-native
route/style/contact/topology decoding, current per-line rendered evidence and
located, hash-bound source/render review. This prevents an aggregate dark-pixel
count or Scene-derived observation from substituting for actual line verification.

Line facts cover ordered straight/polyline/quadratic/cubic/arc routes, closed
paths and filled freeform arrows; both terminals, signed tangents, bends,
start/end markers, physical attachments, native connection IDs, stroke colour/
opacity/caps/joins/dashes/compound bands, crossing/shared/junction decisions and
source-confirmed occlusion. Saved decoding does not import the writer. Unknown
presets/effects and insufficient measurements retain an explicit decision; no
generic diagonal or bbox fallback receives PASS.

The program keeps all obligations while choosing the least expensive sufficient
method. A normal valid PASS stops optional refinement. Only inconclusive bounds
or support select denser finite local measurement. A valid failure cannot be
erased by switching algorithms or widening frozen bounds. Declared-only generic
metrics avoid unrelated SSIM/histogram/edge work. Exact-input raw facts and native
batch work are shared; independent final predicates are recomputed.

Every line has six criteria and current original/render crop references. Clear
cases share readable views across criteria. Non-PASS judgments create persistent
doubts even when no issue text was supplied. Every retained doubt needs enlarged
current evidence and explicit resolution plus the same program criterion's valid
PASS. Long/curved doubts are fully tiled. Measurement/visual uncertainty returns
to evidence review before drawing; established failures use scoped native repair.
Final delivery rejects missing gates, missing line rows/criteria, stale evidence
and wrong attachment bytes.

Independent forward testing exposed normal-flow defects: bent connectors did
not advertise actual corner capability; invalid-proposal deduplication omitted
the current implementation; one incremental template copied an object instead
of its component ID; and route repair copied geometry without its saved frame.
These are repaired with regressions. Invalid component.model types now receive
a located ValueError. Real rendered evidence also exposed shaft/marker overlap
and overly loose pixel bounds. Shaft ownership is corrected; finite cell-union
refinement tightens the same raster target, with complete interior coverage.
Justified higher-resolution rendering is bounded and retains the old evidence.
Program doubts persist as soon as issued and require explicit closure even when
a subsequent measurement passes. No source threshold was widened to pass.
The actual Linux native-arrow repair exposed a further rebind gap: rebuilding
the construction ledger omitted existing line obligations. Rebind now preserves
all doubts and consumed precision history; changed evidence reopens old resolved
questions while preserving their prior resolutions. Independent real-bind tests
verify this, including that empty applicability cannot erase outstanding doubts.
The final counterexample also proves that reordered self-intersecting curved
loops can have identical occupied sets/length; these now require ordered proof
or source-declared simple parts and never receive an insufficient PASS. After
an implementation update, an old visual manifest returns a current review packet
while preserving evidence/issues; it neither crashes next nor reuses old PASS.

Actual renderer calibration retains a material backend difference: PowerPoint
agrees with all 24 frozen bent-preset variants; LibreOffice agrees with 10 and
diverges on 14. The current production per-line render gate rejects all 14 with
valid failure evidence. The raw failed preset report remains FAIL. Encoding the
same frozen routes as editable native explicit paths passes all 24 per-case
native/paint bounds on both backends, without changing the original or 2.5px
calibration bound. This qualifies the representation and rejection behavior,
not universal LibreOffice preset compatibility or automatic rerouting.

Product/core/package/installer version is 1.37.2. Execution policy 1.37.0,
source workflow 1.35.2, design revision 8, Math policy 1.35.3 and runtime/companion
1.14.0 remain independent. Existing bound tasks are not silently migrated. Runtime
binaries/resources and unchanged immutable bundle parts are reused. The active
customized installation, user originals, private tasks/fonts/knowledge, global
skills and published history are preserved.

The evidence is scoped to actual recorded Windows PowerPoint and WSL/Linux
LibreOffice fixtures. It does not accept the lost erroneous artwork, arbitrary
external artwork, physical WPS/macOS or every third-party harness. Artifact hashes
bind identity and coverage; they cannot prove an observer looked or isolate an
editable local host. No provider-token or universal performance claim is made.

lotraAI Image2Slide 1.37.1

lotraAI Image2Slide 1.37.1

Current execution-policy packets now advertise one executable entry: submit.
Their submission_command argument array binds the task and scratch result path;
submission_template binds the immutable template hash and names copy_to. This
replaces obsolete direct component-workflow action hints in observation, binding,
recovery and final audit. Phase operations and required evidence stay in their
existing fields and handlers. Compact packets retain and verify these controls.

This patch changes only packet guidance/template binding and compaction control
fields, plus the requested phase/entry regressions and release metadata. Existing
source, binding, native, layout, font, visual and final acceptance rules remain
unchanged. Direct writers remain guarded. Old bound policies are not relabeled.

Focused tests exercise real observation/binding packet producers, existing
recovery/final handler dispatch through submit, all literal packet kinds,
compaction, changed controls and paths containing spaces. Existing final installed
Windows/Linux workflow and differential/distribution/install gates are retained.
Current counts and exact byte identities are in Packet-submission-checks.json,
Workflow-checks.json, Installation-checks.json and Version-consistency.json.

Product/core/packager/installer labels use 1.37.1. Execution policy remains
workflow-execution-1.37.0; source workflow 1.35.2, design 8, Math policy 1.35.3 and
runtime/companion 1.14.0 stay independent. Runtime Python/binaries/resources are
unchanged; paired repository entries/channels and immutable bundle parts update
differentially. Keep all prior tags/assets, active custom skills, user files and
private tasks/fonts/knowledge. No third-party harness study, complex external-art
acceptance, WPS/macOS host qualification or performance claim is added.

lotraAI Image2Slide 1.37.0

lotraAI Image2Slide 1.37.0

1.37.0 remains a self-contained skill. It makes the registered reconstruction
loop easier to resume and gives its official outputs verifiable completion
criteria. No replacement harness, resident service, account or model key is
introduced. A third-party host with arbitrary code/file access can still ignore
the skill; these are protocol/evidence checks, not hostile-host attestation.

• New tasks use current source/revision/implementation-bound writer grants.
  Product construction and mutations follow the current packet or the existing
  finite action generator. Standalone build requires explicit developer-draft
  mode in private project storage and reports no component/artwork acceptance.
• Original outlines, bounded observation chunks and bounded binding chunks persist
  as immutable checkpoints. Full source/spatial/capability/ownership validators
  seal the complete accumulated data before native construction. This reduces
  request size while retaining whole-source coverage and all final quality gates.
• Packets provide hash-bound submission templates and bounded pending IDs/counts.
  New contexts resume saved work. Explicit prior-bound corrections retain earlier
  snapshots. Source/contract seals commit their checkpoint in the same task-state
  transaction, preventing an interruption from separating the two identities.
• Official export binds the original, complete contract, current independent
  audit, native candidate and final transferred bytes in its private receipt.
  delivery-check and the optional ordinary check_before_finish.py script reject
  intermediates, changed files, empty/incomplete ledgers and stale proofs. The
  caller controls its own reply; no mandatory harness hook is installed.
• An implementation/source/candidate/render-environment change invalidates stale
  ACCEPTED final results in the ordinary next path. Earlier proofs remain in
  history; an old status field cannot substitute for current independent gates.
• SQLite transaction connections close deterministically. Current SKILL.md gives
  one active rule set and progressively links retained source schemas.

Actual Windows PowerPoint and Linux LibreOffice self-authored workflow fixtures
exercise observation through formal two-file export and exact-byte verification.
Negative checks cover bypasses, current-packet/phase restrictions, coverage,
retained corrections, 40 components/200 feature bindings, final-file/receipt
tampering, incomplete ledgers and repeat export. Detailed current counts and
byte-bound results are in Workflow-checks.json, Installation-checks.json,
Backend-and-delivery-checks.json, Final-distribution-checks.json and
Version-consistency.json. These fixtures do not accept external user artwork.
WPS and physical macOS require their own host evidence and remain unqualified.

Product/core/packager/installer/documentation use 1.37.0. Execution policy is
workflow-execution-1.37.0; source workflow 1.35.2, source-features/spatial 1.35.1,
design_revision 8, native Math policy 1.35.3 and runtime/companion 1.14.0 remain
independently versioned. Unchanged runtime executable/resource bytes and immutable
bundle parts are reused. Online installation updates only changed managed files;
source and runtime metadata channels/entries are synchronized incrementally.

Both repositories use dev and numeric 1.37.0 branches, with formal source tag
v1.37.0. Preserve 1.35.4/1.35.5/1.36.0 tags/assets, user originals/deliverables,
private tasks/fonts/knowledge, existing local modifications and global skills.
The active custom local skill is not forcibly updated or relabeled by publication.

lotraAI Image2Slide 1.36.0

lotraAI Image2Slide 1.36.0

Use a capability-qualified local presentation application before the declared
LibreOffice fallback. One common source/geometry/font/OMML generator now emits
Office-compatible native freeform routes rather than custom geometry inside
p:cxnSp. Attached endpoint bindings are never silently discarded. Both saved
structure inspectors check the exact source route and bounds.

Windows PowerPoint and WPS use bounded local COM adapters; the macOS PowerPoint
AppleScript adapter remains experimental until its actual host probe passes.
Readiness requires real PDF export, native PPTX save, edit, save and reopen of
an owned fixture with genuine Office Math. Every final PPTX has its own native
save/reopen gate, independent of the actual PDF and whole-artwork gates. Auto
may use one declared LibreOffice fallback after an actual native PDF failure,
retaining the native failure and naming the real backend. An explicit backend
choice is binding. Existing user files and shared Office sessions are protected.

Core resources are prepared before native qualification. A qualified local
application can satisfy native Math/render/save requirements without installing
LibreOffice. Otherwise the existing independently versioned Math resources are
used, with owned profiles and process-local TEMP/TMP/TMPDIR. A PNG never stands
in for a failed PDF. LO-only export reports target Office as UNVERIFIED.

Users can select a delivery root, with project root as default. Each export
reserves output-slide-YYYY-MM-DD-HHMMSS-mmm, stages on that same volume, verifies
the actual final file, and publishes through an atomic no-replace rename.
Repeated resume uses the same destination; explicit --new-export creates a new
delivery. Only Final_editable.pptx and Report.md are exposed. Task/audit data
remains private in .lotraAI-build and common.safe_path is not widened.

Product/core/packager/installer/current documentation and distributions use
1.36.0. Workflow_version 1.35.2, design_revision 8, the LO Math policy 1.35.3 and
runtime/companion 1.14.0 remain independent. Runtime binary/resource assets are
unchanged. Reuse verified historical bundle parts; upload only the new tail,
skill payload, necessary entries, catalogs and evidence. Source/runtime default
branch remains dev, version branches 1.36.0, formal source tag v1.36.0.

Keep v1.35.4 and v1.35.5 tags/assets, user sources/deliverables, private resources,
the active custom local skill and all global skills. No forced local activation.
Windows PowerPoint and LibreOffice fixtures have actual host evidence. Linux
qualification and final installer/distribution checks are required before release.
Physical WPS/macOS qualification is pending; mocks and WSL do not establish it.
No external user artwork is accepted by these release fixtures. In particular,
the diagnostic repair of a legacy invalid connector changes its old LO display
and is not delivered as an accepted replacement for the user's original.

lotraAI Image2Slide 1.35.5

lotraAI Image2Slide 1.35.5

Repair the complete-bundle download route: the manifest, archive name and result
use the current product version, and verified unchanged parts retain their
original release filenames/URLs. Synchronize the skill packager version with the
coordinator during template generation, so the current source and final skill ZIP
can be built and verified by the supplied packager.

Every iteration now requires a version-consistency gate in the development
AGENTS.md. The maintained gate checks both repositories' current metadata,
README identities, installer pins, catalog records, loose assets, the actual
skill ZIP and complete ZIP including the nested skill. Version-drift regressions
reject stale packager, channel, README and embedded ZIP versions. Shell LF and
Windows PowerShell parser gates still inspect final distributed bytes.

Correct current publication guidance, product tables, 19-part download notes and
historical validation attribution. Stable HTTP tool identifiers no longer carry
an obsolete product version. The runtime Python entry changes only that
diagnostic header; runtime 1.14.0 binaries and existing resource archives do not
change. Workflow_version remains 1.35.2 and design_revision remains 8.

Windows/Linux first/repeat complete-bundle installs, maintained README commands
and full-bundle downloads have separate current receipts in
Installation-checks.json. Final-distribution-checks.json and Version-consistency.json
bind parser/version conclusions to final bytes. Physical macOS qualification,
reconstruction, formula rendering, Office interactive editing and performance
effects are not claimed by this installer patch.

The source and runtime repositories use dev and 1.35.5; the formal source tag is
v1.35.5. Existing v1.35.4 tag and attachments are preserved. Reuse unchanged
resource assets and complete-bundle parts; upload only the changed tail part,
small skill payload, installation entries and current metadata/evidence. No new
runtime binary Release is created. Local user modifications, active runtime,
private data, fonts/knowledge/tasks and global skills are preserved.

lotraAI Image2Slide 1.35.4

lotraAI Image2Slide 1.35.4

Fix the shell bootstrap published in 1.35.3: every shell installation entry uses
LF, and one canonical template renders each public, runtime and skill copy.
Repeated generation preserves identical bytes and modification times instead of
adding nested conditions or repeated parameters. Git attributes explicitly retain
LF for shell scripts and shell templates. The download path has one bounded
official asset-API fallback, with the existing pinned content checks.

The publication gate parses actual loose release attachments, skill ZIP members
and complete-bundle members with Linux sh/bash and Windows PowerShell. It also
checks shell LF bytes, copy/pin identities and unchanged runtime resource records.
Bootstrap regressions exercise download, fallback, HTTP range resume, changed
cache preservation and verified cache reuse on Windows and Linux.

Windows/Linux complete-bundle installation and the public README commands are
recorded separately in Installation-checks.json. Public downloads and final
parser identities are recorded in Final-distribution-checks.json. Physical macOS
installation remains pending; Linux parsing does not qualify a physical Mac.

Product version is 1.35.4; workflow_version remains 1.35.2 and design_revision 8.
Runtime binaries remain 1.14.0. The maintained Python installation core, native
Math implementation and platform resource bytes are unchanged. Runtime repository
changes are limited to shell/PowerShell entry pins, Git attributes, documentation
and the current installation channel. No runtime binary Release is created.
Unchanged immutable bundle parts and existing resource assets are reused.

Preserve global skills, private knowledge/fonts/tasks, source images/PPTs and
historical tags. This is an installation patch: no new artwork acceptance,
Office interactive editing or performance/token claim is added.

lotraAI Image2Slide 1.35.3

lotraAI Image2Slide 1.35.3

The macOS Math backend now runs a fixed Python/UNO worker inside the signed
LibreOffice application, rather than starting its constrained bundled Python
from the project process. Installation startup checks and actual document export
share the same bounded backend. Official application signatures and system
security settings are preserved; workers and new profiles are project-private.

The worker retains native OMML import, native Math formulas, explicit fonts and
sizes, immutable canonical PPTX bytes, per-job input identities and output
receipts. It accepts only startup or export operations. Windows and Linux keep
the external UNO route. Startup probes run afresh on install/repeat; historical
import receipts do not certify the new route. Failures record their phase and
leave unexecuted core checks as NOT_RUN. The old offline compatibility path uses
the same probe instead of its Windows/Linux-only platform expression.

GitHub Release downloads can use the exact official asset API after bounded
failures. Asset names, sizes, content hashes, range responses and owned caches
retain their existing checks. OS bootstrap engine fallbacks use verified fixed
asset IDs. macOS Math DMGs retain the pinned upstream download.

Product version is 1.35.3. The spatial reconstruction policy remains
workflow_version 1.35.2, design_revision 8; old task contracts are preserved.
Runtime binaries remain 1.14.0 and existing Python/OCR/LibreOffice resource bytes
are reused. No new runtime binary release, Setup skill, global install, provider
key, personal data or user-artwork benchmark is introduced.

Focused tests and Windows/WSL formula exports were executed. The hosted route
was actually exercised on Windows, which does not qualify macOS launch/security
behavior. Apple Silicon and Intel physical-host validation remain pending.
See Validation-report.md for the exact scope. This release is a compatibility
implementation with bounded evidence, not a claim of completed Mac acceptance.

Repository defaults/development use dev; numeric version branches use 1.35.3;
the formal product tag is v1.35.3. The maintained installation channel is
activated only after the matching public catalog and assets are verified.