Docs/dead sharepoint links

[XueBin-weston]

Refs #31, and standardises how every robot page offers documents.

Product pages no longer depend on hosts we do not control where the document is ours to publish. 24 documents are now in the file store, up from 2, and every dead link under robot/ is gone.

Page	Dead links
wr65.md	9 → 0
wrl63.md	9 → 0
piper.md	11 → 0
kinova-gen3-lite.md	6 → 0
ranger-mini-v2.md	6 → 0
ranger-mini-v3.md	3 → 0
z1.md (dev-z1.unitree.com)	2 → 0
Google Drive links (go2, g1, b2)	4 → 0
Total	50 → 0

peripheral/sensor/manifold_pocket.md is the only page left with dead links. See Follow-ups.

The rule applied throughout

If the document is ours, publish it. If it is the vendor's, link to them.

A copy of a vendor's manual freezes: our copy of Kinova's Gen3 Lite user guide was r3.1 while Kinova ship r4.0 today, and the Go2 "user guide" on Google Drive turned out to be Unitree's own manual at V1.0 from 2023. A document only we wrote, such as a training deck or a reflash procedure, has no vendor to link to, and an opaque share token is the same failure class as the SharePoint links.

Published to the store

• wr65, wrl63: WR65 and WRL63 are rebadged Realman RM65 and RML63-B. Realman's manual, ROS manual, JSON protocol and interface function description are published for both.
• piper: user manual EN and ZH, CAN protocol reference (Chinese), three STEP models (arm, gripper, drag teaching tool).
• ranger-mini-v3: user manual, CAD, and the charging station manual in EN and ZH.
• go2, g1, b2: our own material: two training decks and the expansion dock reflash guide.

Linked to the vendor

• kinova-gen3-lite: Kinova's resource pages where they exist, so no revision appears in the URL; Artifactory directory listings where they do not.
• All Unitree models: each links Unitree's per-model user manual. As2 has none published yet. The Go2 Drive "user guide" now links to Unitree. Z1 points at Unitree's current developer docs.

Links removed

No source exists for these. The old SharePoint can't be recovered, so the ones that were only linked from it are recorded in TODO.md as lost until found, to be published if a copy turns up.

Removed	Status
wr_arm_sdk SDK manual (wr65, wrl63)	our own document; lost until found
WR65/WRL63 Chinese user manual	lost until found; the store has en only
WR65/WRL63 ROS 2 manual	lost until found; the published integration guide covers ROS 1 only
RM API samples (wr65, wrl63)	lost until found
Ranger Mini 3.0 firmware V6.0.5	lost until found; not on AgileX's site
Piper Chinese quick start	marked "Limited Public Document"; not ours to publish
Piper Windows application	AgileX's own binary; also removed from the bring-up steps
Kinova KORTEX system overview	Kinova publish no Gen3 lite equivalent

Specs corrected against the published manuals

The WR65 and WRL63 pages call the store manuals "the ones to work from", so they now match them:

• Power supply: 24 V rated, 27 V maximum (was 20–30 V).
• Joint limits from each manual's spec table. WR65 J5 is ±180°: the manual says ±128° in its installation section, which looks like a documentation error.
• WRL63 weight 10.2 kg (-B); the page said 10.0 kg.

Ranger Mini 2.0

No longer supported and not being documented, but the dead links are removed. The manual and CAD rows go; the firmware table keeps its rows with the versions unlinked, and points at support for the images.

As2, H2 and NERO unlisted

All three are incomplete. unlisted: true rather than draft, so the pages stay reachable by URL, but they are out of the sidebar and the sitemap. Their product cards on /robot are removed.

docusaurus-lunr-search does not honour the unlisted front matter, so all three routes are also in excludeRoutes. CONTRIBUTING now has an Unlisting a page section with the full steps.

Conventions and tooling

• CONTRIBUTING: publishing now goes through <Downloads> with no hand-written links; naming table uses the real column names and includes zxx; new Unlisting a page section; rationale for config belongs in an ADR, design doc, TODO or LESSONS, not in a comment.
• Rationale moved out of config: docusaurus.config.ts, the sidebars-*.ts files and guides/tags.yml keep only mechanism comments. The reasoning is in ia-proposal.md §9–§11 and release-pages-plan.md §2.
• tutorial → guides: wrfiles.SECTIONS, check-downloads.py, the upload skill, README, the page template and llms.txt. A test now fails if SECTIONS drifts from docusaurus.config.ts.
• Plugin: the staged-file preview now reads zxx and subjects the same way wrfiles.py does, with a test for drift.
• file-hosting.md: status updated now the store is loaded.
• Two worked examples in CONTRIBUTING were wrong: wire-protocol is a subject of api-reference, not a kind, and the display columns show CAD and API reference.

Reviewing it

npm run check:downloads    # all 8 <Downloads> queries resolve
npm run test:files         # 48 tests
npm run build && npm run serve

Serve the production build rather than npm start when checking As2, H2 and NERO. Docusaurus deliberately shows unlisted pages in development.

Every version was read from inside the document, not its filename. That caught SCOUT MINI USER MANUAL 3.0.pdf actually containing V.2.0.0 from 2020, and confirmed RGM3.0.STEP, whose header says RGM2 v2.01, really is the Ranger Mini 3.0 model by measuring it at 722 × 459 × 341 mm against the datasheet's 720 × 500 × 345.

Follow-ups, not in this PR

All tracked in TODO.md:

• Manifold Pocket (first generation): discontinued; the page needs a notice pointing to Pocket2, and its four dead links handled. Grouped in its own TODO section.
• Lost until found: the documents in the table above.
• Dead links inside published PDFs: the Go2 expansion-dock manual and the training decks still link to the old SharePoint from inside the files.
• Automated external link checking (issue Dead links: 53 SharePoint links point at a renamed tenant #31's third criterion): exclude support.unitree.com, which returns 567 to non-browser clients.
• _permcheck/xuebin-20260907.txt in the bucket: needs an AWS admin to delete it.
• Audit 2026-10-07: a read-through of every Markdown file. Items this PR caused are fixed here; the rest are in TODO, starting with questions for the product owner and safety items.
• 8bc8492's commit message is wrong: it adds unlisted: true to h2.md, not a manual link. Left as is rather than force-push during review.

[timothy-weston]

Is the Z1 part of the scope? unitree_z1_dev is not accessible, its a dead link from unitree
Are there future works to be done? If yes, please create or add onto a TODO list to keep track.

[timothy-weston]

Nero has some emtpy resource fields. Should this page also be hidden?

[timothy-weston]

Service manual · Expansion dock itself is fine. There is a deadlink in that file itself, it still points to tangrobot

Take Note: The check of deadlinks should expand to the training slides [G1, B2, etc] as well.

[timothy-weston]

Even though As2 and H2 (potentially to include Nero) are hidden, the Robots intro.md still shows As2 and H2. These should be hidden as well

[XueBin-weston]

All four points addressed, and the branch is now merged up to current main (c4108be), so the conflict is cleared.

1. Z1 scope / dead dev-z1.unitree.com — bdb0b32
The page now points at Unitree's current developer docs (support.unitree.com/home/en/Z1_developer/sdk_intro). The dead dev-z1 host is gone from robot/manipulator/z1.md.

2. NERO empty resource fields — c1f24a9
Hidden. unlisted: true on the page and removed from the sidebar, matching how As2 (348049a) and H2 (8bc8492) were handled.

3. Robots intro still showing As2 / H2 — 531163f
Cards for all three unlisted pages are now dropped.

4. Dead links inside the published files themselves
Added to the TODO as a follow-up. It's a different class of problem: the fix is to the documents, not the pages, and several are outdated enough that they'll need recreating rather than just relinking. This covers both the training decks and the go2 expansion-dock manual you flagged. This PR stays scoped to links on the pages.

@timothy-weston Please help review~ Thanks.

[timothy-weston]

can remove these comments. Claude likes to add comments relating to history/ implementation/ design... its something we need to specifically tell claude not to do, and instead move these kinds of comments and such into ADRs/ changelog/ todo/ etc, instead

[XueBin-weston]

Done in eae2690. The rationale went to ia-proposal §11 and LESSONS, and the rule is now in CONTRIBUTING. The sidebars and guides/tags.yml had the same kind of comments, so those moved too (30f86e5).

[XueBin-weston]

Added since your review: a rule that rationale belongs in ADR/TODO/LESSONS rather than config comments, and the eight comment blocks in docusaurus.config.ts moved accordingly. Three of them were recorded nowhere else and are now an ia-proposal.md §11 amendment. No behaviour change.

[timothy-weston]

Re-checked at eae2690. The config cleanup looks good. The unlisted pages match excludeRoutes exactly (5 of 5), check:downloads and test:files pass, and every new external link resolves. Unitree's support site returns 567 to curl, so I checked Z1 in a browser; it lands on the SDK intro. Four inline comments.

Nit: LESSONS.md:158 says nine comment blocks, but the commit message and the TODO entry say eight. The diff touches nine if the two lunr comments are counted separately.

[timothy-weston]

Good to have the rule here. That's what I was after with the config comments. I'd cut this paragraph down to the rule, the test and the pointer, though. Most of it repeats the reasoning from the LESSONS entry it links to (the public-repo point and the test are both in LESSONS.md:158-159), which is the same rationale-in-the-wrong-place problem the rule is about.

Two specifics:

• "One of the three above" leaves out design docs, but this same commit filed three decisions in ia-proposal.md §11, which was the right place. The list should include the design doc being amended.
• The public-repo sentence is the lesson's reasoning, not something a contributor can act on. Read as a rule, it also conflicts with this commit, which moved "an oversight, not a decision" into ia-proposal.md:321, and that file is public too.

Suggested wording:

Suggested change

**Not in a config file.** `docusaurus.config.ts` and its neighbours record *what* is set, not why it was decided. Leave at most a one-line pointer and file the reasoning in one of the three above. The test is whether the comment would change if the decision were revisited: a note explaining a mechanism a maintainer needs at that line stays, a note explaining a decision moves. This repository is public, so a comment describing the state of our own content — which pages are incomplete, what was an oversight — is published where no reviewer reads it. See [`docs/LESSONS.md`](docs/LESSONS.md), *"Rationale in a config file is filed where nobody looks for a decision"*.

**Not in a config file.** `docusaurus.config.ts`, the sidebars and CI workflows record *what* is set. A comment there may explain a mechanism a maintainer needs at that line. The reasoning behind a decision goes in an ADR, the design doc it amends, `TODO.md` or `docs/LESSONS.md`, with at most a one-line pointer left behind. If the comment would change when the decision is revisited, it moves. See [`docs/LESSONS.md`](docs/LESSONS.md), *"Rationale in a config file is filed where nobody looks for a decision"*.

[XueBin-weston]

Applied your wording in a9c2de2, including the design doc. The LESSONS entry names the design doc as well now, and the eight/nine count is fixed. The §11 wording no longer carries the judgements (30f86e5).

[timothy-weston]

This step still mentions the Windows application, but its row was removed (AgileX's binary, per the description), so the page no longer says where to get it. Either drop the clause or link AgileX's own download page.

Suggested change

	4. **Drive it from code** using `piper_sdk`, or from the Windows application for a first check.
	4. **Drive it from code** using `piper_sdk`.

[XueBin-weston]

Applied your suggestion in 421f8cd.

[timothy-weston]

Three things from the old table aren't in the store, and they're not in the Links removed table in the description either:

• The Chinese manual. The old row was EN · CN; the store has user-manual in en only.
• The ROS 2 manual. wr65-integration-guide-en-v1.0.pdf is the RealMan RM65-B ROS Manual V1.0. It's Melodic/catkin, with no mention of ROS 2 or colcon, so only the ROS 1 half of the old row is covered.
• The RM API samples. There's no api-examples entry for either arm.

Same for wrl63.md. All three were only behind old SharePoint links, so treat them as lost until found: list them in the Links removed table, and record them in TODO as lost until found next to the SDK manual (see the TODO.md comment). If one turns up, for example in the OneDrive copy the English manuals came from, publish it.

Edit: these are now treated as lost until found, our procedure for documents that were only behind the old SharePoint links.

[XueBin-weston]

All three are in TODO as lost until found, with the SDK manual (f6573f9), and in the Links removed table in the description. Same for WRL63.

[timothy-weston]

This PR closes most of this block, but TODO.md still shows it open:

• 104 (Google Drive): done. No drive.google.com links are left in served content.
• 103 (SharePoint rewrite): done except the two on manifold_pocket.md:46-47, which are on me.
• 88 asks for an export from the old SharePoint. That SharePoint can no longer be recovered, so the item can't be done as written; the documents it covers are lost until found. It also still says WR65/WRL63 have "no reachable documentation today", and the section summary at 86 still says the store is empty.

The follow-ups in the PR description aren't in TODO.md either. Can you add them, so they don't only live in a merged PR's description?

• The wr_arm_sdk SDK manual is lost until found. The old SharePoint can't be recovered, so record it in TODO as lost until found, not as a recovery task. Whether to rewrite it or point at the wr_arm_sdk repo instead is a separate decision.
• Delete the _permcheck/xuebin-20260907.txt test object from the westonrobot-files bucket. It's not a repo file. It's a 55-byte permission probe uploaded on 2026-09-07, and its own text reads "permission probe by xuebin, 2026-09-07, safe to delete". It shouldn't still be there: it's publicly served at download.westonrobot.net/_permcheck/…, and every index rebuild reports it, because its key has no <section>/<product>/ part and wrfiles.index_entry() raises ValueError on it. Publishers have no delete permission by design (docs/design/file-hosting.md §6), so the TODO entry should say it needs an AWS admin and how to do it:

  aws s3 rm s3://westonrobot-files/_permcheck/xuebin-20260907.txt
  aws cloudfront create-invalidation --distribution-id E2SQLRWCEUM8UK --paths '/_permcheck/*'
  # done when this returns 403 (what a missing key returns):
  curl -sI https://download.westonrobot.net/_permcheck/xuebin-20260907.txt

• Unitree and the future link checker. This is one clause for the existing link-checking item (TODO.md:385), not a new task. support.unitree.com returns 567 to non-browser clients, so an automated checker would report every Unitree link as broken even though they all work in a browser. Exclude that domain from the checker rather than accepting 567 everywhere.

Edit: clarified that _permcheck/… is an object in the file store bucket, not a file in this repo; updated for the old SharePoint being unrecoverable; documents that were only behind its links are now worded as lost until found, which is our procedure.

[XueBin-weston]

Done in f6573f9: 103 and 104 ticked, 88 closed as can't be done, and the section summary updated. The follow-ups are in TODO: SDK manual lost until found, the _permcheck object with the admin commands, and the Unitree exclusion added to the link-checking item.

[timothy-weston]

Two more for this branch. Neither blocks the merge. Neither changes anything on the site today: no Guides page uses <Downloads> or a staged /_upload/ link, and every CAD file in the store is already tagged zxx. They're latent problems; the first one surfaces the first time someone publishes a document for a Guides page.

The section below is written so you can hand it to Claude as-is. It asks Claude to re-check everything before editing, since these are my findings and your instance may know something mine didn't.

---

Brief for Claude

Context. Repo westonrobot/docs, branch docs/dead-sharepoint-links, PR #44. Everything below was checked at eae2690. If the branch has moved since, find things with the commands rather than the line numbers. Read CONTRIBUTING.md and docs/LESSONS.md first; they set this repo's conventions. Run commands from the repo root.

How to work. Treat each item as a claim to check, not a fact. For each one: run the checks, and apply the change only if they agree with the claim. If a check disagrees, or you see a better fix, don't apply that item. Reply on this PR with what you found and the output that showed it. Don't force-push; the PR is in review.

Item 1: the scripts still list tutorial as a section, but the folder is now guides

The Guides section moved from /tutorial to /guides on 2026-09-11 (docs/design/ia-proposal.md §11). Two lists in scripts/ were not updated:

• SECTIONS in scripts/wrfiles.py. Its comment calls it "the six Docusaurus plugin instances". It decides which store paths are accepted, and publish-files.py uses it (CONTENT_DIRS) to find the pages whose /_upload/ links get rewritten after an upload.
• CONTENT_ROOTS in scripts/check-downloads.py. The folders scanned for leftover /_upload/ links and for <Downloads> queries.

Claimed effect. A document staged under _upload/guides/... is rejected. If someone works around that by staging it under tutorial/, it publishes to a permanent /tutorial/... URL, the Guides page's link isn't rewritten, and neither script reports it.

Checks:

git grep -n tutorial -- scripts CONTRIBUTING.md .claude/skills
#   expect 4 hits: wrfiles.py SECTIONS, check-downloads.py CONTENT_ROOTS,
#   CONTRIBUTING.md <section> row, upload-file/SKILL.md section row
grep -n "path: '" docusaurus.config.ts
#   expect six docs plugins including 'guides'; no 'tutorial'
ls -d tutorial guides
#   expect: tutorial does not exist
grep -n "CONTENT_DIRS" scripts/publish-files.py
#   expect CONTENT_DIRS = wrfiles.SECTIONS, used in pages_referencing()
python3 -c 'import sys; sys.path.insert(0, "scripts"); import wrfiles; wrfiles.key_from_upload_path("_upload/guides/foo/foo-user-manual-en-v1.pdf")'
#   expect NameError_: section must be one of (... 'tutorial' ...), got 'guides'
curl -s https://download.westonrobot.net/index.json | python3 -c 'import json,sys,collections; print(collections.Counter(f["key"].split("/")[0] for f in json.load(sys.stdin)["files"]))'
#   expect only 'robot'. If any key starts with 'tutorial/', STOP and ask: published keys are permanent.

Suggested change. Adapt it if you see something better.

1. wrfiles.py: change "tutorial" to "guides" in SECTIONS.
2. check-downloads.py: replace its own tuple with CONTENT_ROOTS = wrfiles.SECTIONS. Import wrfiles the way publish-files.py does (sys.path.insert(...) then import wrfiles # noqa: E402), so there is one list instead of two.
3. Update the two docs that repeat the list: the <section> row of the naming table in CONTRIBUTING.md, and the section row in .claude/skills/upload-file/SKILL.md.
4. Add a test to scripts/test_wrfiles.py that fails if SECTIONS stops matching the path: entries in docusaurus.config.ts, and also fails if check-downloads.py scans a different set. There's a sketch below. Checked at eae2690: it passes with the fix, and fails when either list is put back to tutorial.

Do not change. These mention tutorial deliberately:

• the 24 from: '/tutorial/...' redirects in docusaurus.config.ts, which keep old customer links working;
• tutorialSidebar in sidebars-guides.ts, which is a name in code, not a path;
• docs/adr/0002-…, which is a dated record.

Test sketch

Add import importlib.util, import pathlib and import re to the imports, and REPO = pathlib.Path(__file__).resolve().parent.parent below import wrfiles as w. Then:

def docs_plugin_paths():
    config = (REPO / "docusaurus.config.ts").read_text(encoding="utf-8")
    return set(re.findall(r"^\s*path:\s*'([^']+)',", config, re.M))


class SectionsAreTheSiteSections(unittest.TestCase):
    """`SECTIONS` must equal the `path` of every docs plugin instance in
    `docusaurus.config.ts`, and `check-downloads.py` must scan the same set."""

    def test_sections_match_the_config(self):
        paths = docs_plugin_paths()
        self.assertTrue(paths, "no `path:` entries found in docusaurus.config.ts")
        self.assertEqual(set(w.SECTIONS), paths)

    def test_check_downloads_scans_the_sections(self):
        spec = importlib.util.spec_from_file_location(
            "check_downloads", REPO / "scripts" / "check-downloads.py")
        check_downloads = importlib.util.module_from_spec(spec)
        spec.loader.exec_module(check_downloads)
        self.assertEqual(set(check_downloads.CONTENT_ROOTS), set(w.SECTIONS))

Item 2: CONTRIBUTING.md's language rules leave out zxx

wrfiles.LANGS accepts zxx, which means "no language" and is used for CAD, firmware and wiring diagrams. The wrfiles.py error message tells people a CAD model is zxx, and the example filenames in CONTRIBUTING.md use it. But the two places that state the rule list only en, zh, zh-hans, zh-hant. Nothing in the code requires zxx for CAD, so someone following the table would publish a CAD file tagged en.

Checks:

grep -n "^LANGS" scripts/wrfiles.py   # expect zxx in the tuple
grep -n "zh-hant" CONTRIBUTING.md      # expect the <lang> row and the "Chinese is zh" note; neither mentions zxx
grep -n "zxx" CONTRIBUTING.md          # expect only example filenames

Suggested change: add zxx to both places and say what it's for. The wording is up to you.

When you're done

npm run test:files        # all pass, including the new tests
npm run check:downloads   # expect "All 8 <Downloads> quer(y/ies) resolve." (more if pages were added)
git grep -n tutorial -- scripts CONTRIBUTING.md .claude/skills   # expect no hits

Do the break-and-restore check on the new test: put "tutorial" back in SECTIONS, confirm the test fails, restore it, and confirm it passes. Commit in the repo's usual style, then reply here with what you applied, anything you did differently and why, and anything you disagreed with.

Already open on this PR, from my earlier review. These are pointers only; the details are in each thread. Handle them the same way: check, then apply or reply in that thread.

• CONTRIBUTING.md:244: trim the new "Not in a config file" paragraph
• piper.md:39: the bring-up steps still mention the Windows application
• wr65.md:55: three WR65/WRL63 documents not accounted for
• TODO.md:103: finished items still unchecked; follow-ups not tracked
• the summary of the earlier review: LESSONS.md says nine comment blocks, the commit and TODO.md say eight

[timothy-weston]

The same procedure applies here. The two tangrobot.sharepoint.com manuals on manifold_pocket.md (MindCloud GO manual and MindCloud manual) are lost until found, like the other documents that were only on the old SharePoint. This item still says the SharePoint pair "is covered by the rewrite item under File hosting above", and that rewrite item assumed they could be exported and rehosted. Can you reword it to record both manuals as lost until found?

The two app links on www.manifoldtech.cloud (the MindCloud GO APK and MindCloud for Windows) are best effort. The Manifold Pocket is no longer in production, so the vendor may never renew that certificate, and any vendor link may stop working. The item's "a lapsed certificate may come back on its own" shouldn't be counted on; as of today the certificate is still expired (notAfter 2026-07-01). Can you word that part as best effort too?

On best-effort basis. The PR description says Manifold's download page (version.manifoldtech.cn/download/mcs?lang=en) hosts the manuals as well as the software.

Edit: added the manifoldtech.cloud app links as best effort, since the product is no longer in production.

[timothy-weston]

Since Pocket 1 is no longer in production, likely need state that the product and support for this is not longer available and redirect to pocket 2...

[timothy-weston]

Second thought, maybe better to hide pocket 1. Leave this portion to me, once you've rectified the others, should be okay to merge. I'll create another branch to fix this

[XueBin-weston]

Reworded in f6573f9: manuals lost until found, app links best effort. In cf96220 I moved everything for the Pocket into its own TODO section, Manifold Pocket (first generation), including the discontinued notice and pointing readers to Pocket2. I left the page itself to you.

[XueBin-weston]

@timothy-weston Thanks for the review. Everything from 5–7 Oct is addressed, and the branch is pushed. Summary below.

Your review threads

• Config comments: rationale moved out of docusaurus.config.ts (eae2690), and the CONTRIBUTING paragraph cut to your wording (a9c2de2).
• PiPER: Windows app removed from the bring-up steps (421f8cd).
• WR65/WRL63: the SDK manual, Chinese manual, ROS 2 manual and RM API samples are now in TODO as lost until found. Finished items are ticked and the follow-ups are tracked there (f6573f9).
• Manifold Pocket: I haven't touched the page. All its open items are in one TODO section, Manifold Pocket (first generation), including the discontinued notice and pointing readers to Pocket2 (cf96220).

Your brief

• Item 1: SECTIONS and check-downloads.py now use guides. There's a test that fails if they drift from docusaurus.config.ts; I checked that it fails with tutorial and passes again once restored (5afb33f).
• Item 2: zxx added to CONTRIBUTING (155d4f5).

Audit of all .md files
I went through every Markdown file. Anything this PR caused or touched is fixed here:

• WR65/WRL63 specs now match the manuals we publish: 24 V rated / 27 V max, joint limits, WRL63 10.2 kg (8a37e9a). WR65 J5 is ±180°: the manual also says ±128° in one place, which looks like a documentation error.
• Pointers this PR left stale: PiPER CAN reference marked as Chinese, A2 → As2 link removed, Ranger V3 calibration note (a10f823).
• CONTRIBUTING now describes publishing through <Downloads>, plus unlisting steps (abc8be5, 55e27f9). The design doc and upload skill match it (79d6e6f).
• Leftover tutorial paths (cc9f785), file store status (7d3ad91), sidebar and tag rationale moved to the design docs (30f86e5).
• Plugin preview now reads zxx and subjects, with a drift test (57ad362).

Everything else is in TODO under Audit 2026-10-07 (9e0ff83, 0a31a70). It starts with questions for the product owner and safety items. Three I'd flag:

• publish-files.py --publish can overwrite a published document, and treats an unreadable store as empty.
• Deploy only runs on push, so publishing or retiring a document doesn't update the live tables until the next merge.
• TODO.md is public and holds IAM/MFA details that should move to the private infra repo.

Open points

• WRL63 force-sensing weight (10.1 kg) has no source and now reads lighter than the standard arm. Recorded in TODO.
• The published WRL63 manual contains RM65 text. Needs raising with Realman.
• The records no longer point to the old OneDrive; the store is where documents live now.

Checks: npm run test:files (48 pass), npm run check:downloads (all 8 resolve), build clean.