v0.24.0 — Portable research workbench and source-bound writing

v0.24.0 — Portable research workbench and source-bound report writing

This release connects selected research state, author baselines and local figure scripts across
folders and tool-using hosts. CLI/Python and the optional local MCP server use the same business
functions. The package supplies scientific tools and current context; the host develops the argument.

Complete task paths

• Export, inspect and restore a selected workbench, including its current sources, author companion,
  baseline and local scripts. Internal declarations become relative; explicitly mapped external
  result roots relocate without rewriting author prose or script strings.
• Use nine task-sized functions for material inspection, analysis, saved reports, surface research,
  writing, manuscript export, author return and workbench movement. No model API is mandatory.
• Select exact JSON numbers from native reports, named method/setup records and supplied reference
  files. Prepare bound Methods/Results/Discussion and a native Results table without manufacturing
  a general topic proposal or copying numbers into a second evidence list.
• Compare defined scalars using differences, ratios, relative changes or reductions. The reference
  denominator, both parent supports and located method definition travel with each comparison.
• Supply source-linked structured formulas once and use their math tokens in the draft. Word keeps
  native editable fractions, subscripts and superscripts; author-local expressions take priority.
• Continue from retained current reports and author drafts. Recalculate affected values and table
  cells while keeping wording; source-definition or interpretation changes still need host judgment.

Installed examples

Install the matching wheel with [docs] for Word output or [docs,mcp] for a local MCP host.
Run the public scripts from any writable directory using absolute script/output paths:

python /path/to/report-writing/run_example.py /path/to/new-report-output
python /path/to/author-collaboration/run_example.py /path/to/new-author-output
python /path/to/portable-workbench/run_example.py /path/to/new-author-output /path/to/new-portable-output

The report example uses constructed records and tests a denominator/source refresh plus native
math. The portable example moves an analytical author manuscript and preserves its current values,
wording and comments. These runnable examples demonstrate installation and collaboration behavior,
not independent CFD validation. See the host guide for real-task use.

Evaluation and release evidence

Known-project pressure-report and conjugate-heat-transfer paths use existing mature cooling results.
Actual second-host execution uses Claude Code configured with DeepSeek, not an official Claude model; connection,
tool execution, first writing and feedback-assisted revision are evaluated separately. Earlier
report trials exposed denominator and interpretation failures rather than being recorded as passes.
The single follow-up completed preparation, bound writing and native Word export through the new
interface, with a CLI-reported cost of USD 2.446 under its USD 3 setting. All 64 selected/derived
values matched the unchanged source records, relative percentages used their declared reference
denominators, and native fractions/superscripts rendered correctly in a two-page candidate.
This numerical agreement does not mean the first draft passed scientific review.
It still called one static-pressure difference a loss and inadequately named one mixed-weight
ratio. These are residual interpretation errors, not accepted scientific findings. The first
candidate is retained unchanged; the result supports the tool path, not autonomous scientific writing.

The run also exposed an oversized full-report read and an unnecessary question-ID coverage
requirement. A compact writing view and evidence duties that distinguish a question from a finding
address those concrete product problems; they are post-trial fixes, not another host retry.

The six Windows/Linux Python 3.10–3.12 test jobs and installed-package smoke check passed.
The wheel and source distribution below are the artifacts of the verified PR build; the merge
has the same source tree. An isolated wheel replay outside the checkout also passed.
Verified PR CI and
push CI; source: 1c17c383f0b60b2d176a218bf01f4291d7bebaf3.

Retained scope

The flow and heat-transfer tasks are different paths in a known project, not two unfamiliar domains.
Suitable mature real transient or multiphase CFD materials remain unavailable. This third-class
evaluation stays not yet evaluated; saved-time operators and battery control-model regressions
do not replace it. This version does not establish broad scientific generality or submission readiness.

Selective workbench transfer is not whole-project SQLite migration. Native solver archives remain
external; their verified extraction scope is unchanged. A matching support-domain declaration is not
proof of physical equivalence or causality. Formula trees provide native notation, not a new formula
solver. Structured comparisons and automatic token refresh cannot revise scientific interpretations.

Scientific interpretations still require author and host review.