Separation arc — one workplan for GOAL 1

Epic bean: folio-assistant-7x5n (under milestone vuip). Issue: #1770. See the epic bean. This page consolidates five stalled sessions’ handovers into one arc with nine stories. The stories own no new code. Each one points at the beans that already exist. Where two of those beans describe the same work, the story names which one survives.

Why one arc: five sessions each left a handover that said “pick up next: …”. Read side by side, the lists overlap, contradict each other in two places, and share one cause of stalling: the merge treadmill. Main moves about every 3 minutes, and a merge, regen and CI cycle takes about 22 minutes. An arc that ignores that stalls again the same way.

Status bars (update on every story transition)

Measured 2026-10-01 from beans list --json, recursive over each parent, scrapped excluded.

GOAL 1  vuip  repo separation          ██████████░░░░░░░░░░  74/143  (27 in-progress, 42 todo)
  vke6  layer split (#223)             █████████████░░░░░░░  52/79
  iirv  cat-harness / -tools split     ███░░░░░░░░░░░░░░░░░   3/22
  9umr  concern subgraphs              ░░░░░░░░░░░░░░░░░░░░   0/2   (PR0, PR1 merged; beans not closed)
  w2gr  cat-harness-tools (server)     ████████░░░░░░░░░░░░  steps 1–3a merged, 3b open
  fnx4  KG subscriptions               ████████████░░░░░░░░  slices 1–4,7,8 merged; 5+6 in #1756
  uhkv  SMART stack (sibling: n3ni)    ████████░░░░░░░░░░░░   5/13

ARC 7x5n stories (15:15 UTC)
  S0 main green          ██████████  DONE: #1774 → a249bd3 (+ #1769 restored PR0a and pins)
  S1 bookkeeping         █████████░  hx65 closed; 7dek / ejye / ybwt waiting on stated evidence (ybwt → 63wl)
  S2 merge treadmill     ▒▒▒▒░░░░░░  #1754 in progress
  S3 drain PRs           ██████░░░░  R1 done (#1769 + guard #1785); #1756, #1735, #1753 remain
  S4 direction/placement ███████░░░  #1776 merged (R5 26→0, R6 →0); audit #1778 merged; S4-b #1787 in progress; syiq, lthi queued
  S5 code → tools        ██░░░░░░░░  C1 #1786 merged; 70lx, 8lcl, y9r6, vj2p, saqd, bbza queued
  S6 standalone          ░░░░░░░░░░
  S7 seed repos          █████░░░░░  all 5 target repos exist and are empty
  S8 cut over            ░░░░░░░░░░
  UPSTREAM               ██████████  bootstrap#1 f75a216, bootstrap-tools#4 3046412; pin bump = bean pw9j
main CI                                GREEN

Legend for the story row: ░░ todo, ▒▒ in progress, ██ done, !! blocked on the owner.

Stories

#	story	absorbs (existing beans / PRs)	blocks	delegable?
S0 `hx65`	Main green. Fix the five red jobs on `cdb0a018`. Bump the submodule pins (bootstrap `ebfa406`, bootstrap-tools `03832a8`).	handover fix-ups; `14ve`	everything	yes, one agent
S1 `a4of`	Bookkeeping truth. Close beans whose work is merged, and reconcile `fnx4`’s boxes with #1721.	`y5si`, `ybp4`, `ejye`, `ybwt`, `7dek`, `xsqm` (already done)	S4, S5	yes, here
S2 `0mf0`	Merge-treadmill fix. Land `merge:main` (#1754), then run regen in CI on the merge result (`d33q` part B).	#1754, `d33q`, `oz5w`	S3–S8 throughput	yes, one agent
S3 `ga6u`	Drain in-flight PRs. #1756 (fnx4 5+6), #1747 (cmsl 3), #1735 (qsx4, owner judgement), #1753, #1581 (stale plan; close or rebase).	those PRs	S5	yes, one agent per PR
S4 `rfuq`	Direction and placement. No upward imports, and content goes to its owner.	`zlmp`, `yj6r`, `zhg2`, `r3gy`, `2j2r`, iirv `pzwb` `63wl` `tlat` `4fv8` `apcg` `8fq9` `f8wp` `p9bu`, 9umr finale	S5	yes, one agent per placement PR, serialised
S5 `txue`	Code out of cat-harness. All code goes to `cat-harness-tools` (D1).	`w2gr` 3b ≡ iirv `70lx` (one story; `70lx` survives, `w2gr` becomes its child), `8lcl`, `y9r6`, `vj2p`	S6	yes, but one agent at a time (touches every import)
S6 `ybsz`	Standalone rehearsal. Each staged instance builds and passes from a fresh sibling clone.	iirv `ho66`, `pyds`; vke6 `izqr`, `mer2` → `tndo` → `zmdo`, `wggr`	S7	yes
S7 `mgxw`	Seed the staging repos (sibling strategy). One seed PR per repo, on a branch, never to `main` of the target until approval.	iirv `smbc` (owner), `iai8`, `w1gy`; `n3ni` (sibling, smart-*)	S8	seeding yes; authorisation is the owner’s
S8 `w0at`	Cut over. folio-assistant consumes each instance from its repo: a submodule for code it imports, a subscription for KG content. The in-tree copy is deleted only with your OK.	iirv `syzb`; `fnx4` (subscription path); `4475`	—	owner gate

The sibling strategy, adopted here

The sibling working on n3ni (smart-* → litlfred/smart-*) follows this sequence: (1) push generic code down first (into fhir-harness), (2) seed the staged directory into its litlfred/<name> fork on a branch, one PR per fork, (3) folio-assistant subscribes, (4) cut over only on your OK. Then, later, the repo moves to WorldHealthOrganization/* once everything works.

This arc uses the same four steps for the non-SMART instances. The one refinement is in step 3. An instance that folio-assistant imports code from (cat-harness-tools, bootstrap-tools) has to be a submodule, because a subscription materialises KG content and not a module graph. An instance whose content is only read (cat-harness’s skills, who-iris, large-datasets) can be a subscription, which is fnx4’s path. Both cases keep step 4 behind your gate.

Gap analysis

Each gap is something that no existing bean or PR covers, or a place where two of them disagree. Gaps are ordered by what they block.

#	gap	evidence	consequence if left	goes to
G1	Main is red, and every handover deferred it.	`Code-quality gates` at `cdb0a018`: failing steps are workflow skill refs (Repository gates #13), bun test, playwright test, registration chain current. `Docs site`: Regenerate skill instruction pages. Four successive merges went in without CI finishing.	Every delegated agent inherits red, can’t tell its own breakage from the baseline, and stalls.	S0
G2	Two beans, one move. `w2gr` 3b (“move the cat-harness/src server modules into cat-harness-tools”) and iirv `70lx` (“git mv the unambiguous code”) are the same `git mv`, planned by two sessions. The iirv handover already saw `cat-harness-tools/` appear on main from the other session.	Handovers 2 and 4.	Two agents do the move twice, and the merge conflict is the whole tree.	S5: `70lx` survives, `w2gr` closes into it
G3	Merged work still reads as open. `y5si` has every box ticked. `ybp4` lacks only “CI green”, which it got (#1687). `ejye`, `ybwt` and `7dek` are merged (#1758, #1760). `xsqm` is closed but the handover lists it as blocking. `fnx4`’s boxes don’t reflect #1721.	Bean files on main.	A new agent claims done work, as happened with #1690.	S1
G4	The treadmill is unaddressed by any bean that has an owner. `d33q` part B (regen in CI on the merge result) is the systemic fix. #1754 holds part A and is unmerged.	Three handovers measure the cost: 8+ rounds; “a `dirty` PR cannot be merged at all”.	Every story below S2 pays about 22 minutes per main movement.	S2
G5	Two cutover mechanisms, no rule for which applies. iirv’s plan ends in submodules (stage 6). `fnx4` and the sibling’s `n3ni` end in subscriptions.	The plans side by side.	The two staging repos get cut over two different ways for no stated reason.	S7/S8: the rule above (import → submodule, read → subscription)
G6	~~Some target repos don’t exist~~ Resolved 2026-10-01: the owner created `litlfred/{folio-assistant-core, folio-assistant-sci, fhir-harness}`, all three verified empty. `large-datasets` and `agent-skills` fold into cat-harness.	`git ls-remote`, 2026-10-01.	—	S7
G7	~~`litlfred/cat-harness` and `-tools` already hold commits~~ Resolved 2026-10-01: both are empty (fresh clone, 0 commits on `main`), so a seed is the first commit and no read-first is needed.	Clone of each.	—	S7
G8	Orphan sidecars need a deletion ruling. 8 detangle sidecars and 1 kg-qa sidecar are orphaned by moves, and `kg:detangle:check` fails on them.	iirv handover.	A gate stays red, or an agent deletes them unasked.	Your decision (S0)
G9	The submodule pins regress on merge. Main twice put bootstrap and bootstrap-tools back to older commits, and README writers are duplicated here.	iirv handover.	bootstrap’s “generated by” notices get dropped, and a merge silently de-inits submodules.	S0 (bump) + S4 (delegate README writers)
G10	`readme:subgraphs:check` red on `02f16ae98bb` is unexplained. The log redirects to Azure blob, which this proxy denies.	Handover 3.	The gate may recur with no diagnosis.	S0. `gh api …/jobs` works from here, so retry the read
G11	`fnx4` slice 7 layout collides with staged instances (`<root>/<harness>/` vs `who-iris/`).	Handover 5.	The first real subscription (`litlfred/ihris`) writes over a staged directory.	S8 prerequisite
G12	GOAL 1’s own falsifier has no definition. `zmdo` waits on `tndo` (MVP undefined), which waits on `mer2` (split `folio_init`). You ruled “split the operation”; it is not built.	Beans.	Nobody can say whether GOAL 1 is done.	S6
G13	The specialised agent types can’t reach GitHub. `ci-health-watcher` spawned in this session had no GitHub MCP tools and an invalid `GH_TOKEN`.	This session, 2026-10-01.	Delegated diagnosis comes back “could not check”.	Delegation rule below
G14	~~IRIS-specific code has no home~~ Ruled 2026-10-01: the 5 files stay in `who-iris`, and the finding becomes a QA warning, not a failure (“ok b/c small # tools”). No `who-iris-tools` repo.	`eayu`.	—	S0 (severity change)

Migration plan — checklist for status-bar monitoring

Each line is one observable fact. Tick it only on evidence: a merged PR sha, a green run id, or a bean id moved to completed. The status bars above count these ticks.

S0 `hx65` — main green

the five failing jobs on main, each named with job, step and cause
fix PR merged on per-job verified green (not “merged before CI”)
submodule pins: bootstrap ebfa406, bootstrap-tools 03832a8, translation templates re-extracted
readme:subgraphs cause read from the job log (G10)
your ruling on the 8 + 1 orphan sidecars recorded on the bean (G8): widen the scan, keep the files

S1 `a4of` — bookkeeping truth

y5si, ybp4 completed with PR evidence
ejye, ybwt, 7dek completed (#1758, #1760)
fnx4 boxes ticked for slices merged in #1721
w2gr re-parented as a child of 70lx’s story, and its 3b list copied there (G2)

S2 `0mf0` — treadmill

#1754 merge:main merged
d33q part B: CI regenerates on the merge result and pushes a fix-up commit, so a branch no longer races main
measured: median cycle from “PR green” to “merged” before and after

S3 `ga6u` — drain in-flight PRs

#1756 merged (fnx4 5+6)
#1747 merged (cmsl step 3)
#1753 merged (i2kp)
#1581 closed as superseded, or rebased (the r0tm plan predates #1687)
#1735: your per-entry ruling on dependents → subgraph: true, then merged

S4 `rfuq` — direction and placement

check:import-direction + check:reference-direction at 0 wrong-direction edges (zlmp, yj6r, zhg2, r3gy)
2j2r: ~44 readers of decl.directories follow the checkout
owner ruling 2026-10-01: large-datasets becomes a subgraph of cat-harness. Its one upward code import (scripts/gen-id-lookup.ts → folio-assistant-core/schemas/catalogue.js) and its needs: folio-assistant-core go first.
owner ruling 2026-10-01: agent-skills (16 library sources + voices) becomes a library subgraph of cat-harness. It has 0 code imports from core, and its needs: folio-assistant-core plus 2 prose mentions go first.
placement PR2 pzwb · PR3 63wl · PR4 4fv8 · PR5 tlat · PR6 apcg · PR7 8fq9 · PR8 f8wp · PR9 p9bu
9umr finale: last five tool skills out of folio-core

S5 `txue` — code out of cat-harness

1a 70lx (absorbs w2gr 3b): server, tools, routes, rbac, auth, mcp → cat-harness-tools/; MCP server starts and lists the same tools
1b 8lcl: Zod → tools; generated JSON Schema stays
1c y9r6: cat-harness content holds no .ts (D2)
1d vj2p: each instance hosts its own generated outputs (D3)

S6 `ybsz` — standalone rehearsal

ho66 check:cat-harness-standalone green from a fresh sibling clone
mer2 instance-init split; tndo MVP written down; izqr bootstrap on an empty repo
wggr stub inversion (<stub>/docs)

S7 `mgxw` — seed staging repos (sibling strategy)

current contents of litlfred/cat-harness and litlfred/cat-harness-tools read and recorded (G7): both empty, 2026-10-01
your authorisation (smbc)
seed PR on a branch in each target repo; fresh-clone QA (w1gy)
empty repos created by you for core / sci / fhir-harness (G6): created 2026-10-01, verified empty via git ls-remote

S8 `w0at` — cut over

import-consumed instances are submodules at the same path (syzb)
read-consumed instances are subscriptions (fnx4; G11 layout fixed first)
in-tree copy removed on your OK only, per instance

Delegation

What this session can hand to a sub-agent, and the rules that keep it from stalling the way the last five did.

Agent type: use general-purpose (all tools) for anything that touches GitHub. The specialised types run without the GitHub tools here (G13).
Isolation: one git worktree per agent, one branch, and one PR opened at the first commit.
Collision rule: S5 and the S4 placement PRs each touch a large share of the tree. Run at most one of them at a time. S0, S2 and the S3 PRs are disjoint and run in parallel.
Claim first: bun run cat beans:claim <id> before any edit. The claim commits to main, so keep both sides at the merge.
The merge cycle every agent runs: merge main, git submodule update --init --recursive, bun run cat regen until it reports 0, bun run cat gates, push. Run state:visualizer last.
Merging: owner ruling 2026-10-01: “yes you may merge green PRs”. An agent merges once every job on its current head is green and the PR is mergeable. Not before. Merging before CI finishes is not authorised; earlier sessions did that on an instruction specific to them.
Report shape: the agent ends with a table of the checklist lines it ticked, each with its evidence (sha, run id, bean id), and the lines it could not tick, each with the reason.

Seed readiness — cat-harness and cat-harness-tools (measured 2026-10-01, main `cdb0a018`)

Method: git ls-files and a relative-import and markdown-link scan over every tracked .ts and .md file, excluding test/results/. Edges are counted per file-to-target reference and grouped by subgraph.

cat-harness  (KG repo)       ███░░░░░░░░░░░░░░░░░  not seedable yet
cat-harness-tools (code)     █░░░░░░░░░░░░░░░░░░░  not seedable yet (C1 ruled: tools below core)

#	blocker	measure	story
R1	cat-harness declares 20 directories whose path is another instance’s (`who-iris/library`, `folio-assistant-core/skills`, `smart-base/methodologies`, …)	20 of 45 declarations	cmsl step 3, #1747 (S3)
R2	code still in cat-harness	1,522 `.ts` files: scripts 816, schemas 208, src 76, test 58, plus the 363 below	S5 `70lx`, `8lcl`
R3	`.ts` inside KG subgraphs (D2 → JSON)	363: content 337 (block manifests), skills 12, translations 7, tools 5, types 2	S5 `y9r6`
R4	KG → code edges the cut severs	412, of which 363 are `content → schemas` (block manifests importing Zod builders). The rest are tools→schemas 11, skills→schemas 13, translations 7	S5 `y9r6` (data, not imports)
R5	upward links from cat-harness	26 (all markdown): docs→who-style-guide 9, docs→sci 8, docs→core 5, skills→sci/who-iris/beans 4	S4
R6	references to the repo root	13 (5 ts in scripts, 1 in schemas, 7 md)	S4
R7	main is red	5 jobs	S0

What is already clean: 0 references from cat-harness into cat-harness-tools. 231 code→KG edges (tools reading the KG is the allowed direction). 391 inbound edges from higher instances (pointing down is allowed).

C1, a layering conflict, ruled 2026-10-01: tools go BELOW core. cat-harness-tools needs only cat-harness (and bootstrap-tools), and folio-assistant-core may depend on it. The MCP-server and tool-implementation parts that need core move up into core. Ruling 2 now reads: core must not depend on the MCP server. Layer order: bootstrap → bootstrap-tools → cat-harness → cat-harness-tools → folio-assistant-core → {sci, fhir-harness} → smart-*. The text below records the conflict as it was measured.

cat-harness-tools declares needs: folio-assistant-core, and ruling 2 (2026-10-01) says core must not depend on cat-harness-tools. But 88 references from folio-assistant-core point into cat-harness code (scripts, src, schemas), and more come from sci (11), who-iris (12), smart-trust (7) and fhir-harness (6). Under D1 that code moves to cat-harness-tools. Core would then need tools while tools needs core, which is a cycle.

Dense subgraph clusters (where the edges are)

Inside cat-harness, the heaviest edges between subgraphs:

edges	from → to	reading
866	scripts → schemas	the code core. Both move to tools together (Zod), so this is not a cut
363	content → schemas	the main cut: block manifests import Zod builders. D2 makes them JSON
201	scripts → content	code reading KG content. Allowed after the split
155	scripts → src	code ↔ code, which moves together
88	docs → skills	KG ↔ KG, stays together
48	docs → processes	KG ↔ KG
46 / 45	test → schemas, src → schemas	code ↔ code

Skills by concern (files): sdlc 63 · authoring 42 · kg 35 · ui 28 · process 24 · library 23 · conduct 18 · requirements 7 · folio-core 6 (the last five go out in the 9umr finale).

Largest subgraphs by file count: library 2,963 (29 entries; JSON-LD 1.1 501, arXiv 2607.20636 462) · docs 1,487 · test 1,028 · scripts 996 · qa results 966 · content 783 · translations 600 · uml 291.

The per-item placement verdicts (every skill, tool, scenario, process, role) are in placement-audit-2026-10-01.md, which an agent is producing now.

MCP placement (owner rulings 2026-10-01, refined the same day)

MCP gets a dedicated subgraph inside the tools concern group, split by layer:

cat-harness/skills/tools/mcp/ holds generic MCP: the basic mapping from skills to tools, kept consistent, plus contract, projection and assembly.
folio-assistant-core/skills/tools/mcp/ holds folio-specific MCP surfaces: folio tools, content-adapter tool registrars, and the adapter-scaffold half of folio_init (mer2).

First pass, to be verified per item by the placement audit: all 5 MCP and tool skills read as generic (0–8 folio mentions against 24–136 skill or tool mentions). The folio-specific candidates are among the implementations: folio-init, readme-sync/readme-audit, preferences, degradation, translation. The registry tools/index.ts is likely a SPLIT, with the generic registry in cat-harness and the folio tool list in core. Note that “folio” also matches the repo name, so verdicts are made on subject, not on grep count.

piece	target
`mcp-assembly`, `mcp-contract`, `mcp-projection` (now in `skills/folio-core/`)	`cat-harness/skills/tools/mcp/`
`skills-and-tools`, `covered-is-not-reachable`	`cat-harness/skills/tools/`
tool definitions `cat-harness/tools/*.ts`	stay, converted to JSON (D2)
`cat-harness/src/tools/*` (20)	`cat-harness-tools`
`cat-harness-tools/adapters/mcp-server` (16)	stays; any part that needs core moves to core (C1)
MCP spec, `agent-skills/library/mcp-2026-specification-2026-07-28`	cat-harness library (agent-skills ruling)

Tracked on 9umr (finale) under S4.

Placement audit (#1778) and the owner rulings on its questions (2026-10-01)

The audit checked 2,266 items (skills, voices, tools, processes, decisions, roles, stories, actors, capabilities, methodologies, code lists, templates, schemas, library entries, script and code groups):

PLACEMENT    ████████░░░░░░░░░░░░  945 OK / 2266
  CODE    → cat-harness-tools   790   (S5: 70lx 553+, 8lcl 181)
  MOVE    → owning instance     295   (placement PR2–PR8 + 160 unplanned)
  TO-JSON (D2)                  127   (S5: y9r6)
  SPLIT                          89
  AMBIGUOUS                      20   → 7 questions, all ruled below

Q	question	owner ruling
1	integration-watcher family (compute, detangler, devils-advocate, integration-watch, narrative-asserts-code)	move the family to sci; the generic `integration-watcher` stays in cat-harness as their base
2	editorial graph (content-graph, uses-editorial-review, corpus-grep)	split: the `uses[]` relation stays in cat-harness; the Lean comparison moves to sci
3	milnor-exposition-standard, markdown-render-check	Milnor to sci; the render check splits (generic markdown stays, LaTeX/maths to sci)
4	10 pipeline modules with maths vocabulary (qa-criteria-registry, lean-lexer, render-latex, …)	split in S5: the generic pipeline goes to cat-harness-tools; the maths checkers go to sci’s code and plug into it
5	shape of large-datasets and agent-skills inside cat-harness	dissolve into the concern groups (matches PR8 3A)
6	`skills/authoring/` vs the `content` concern code	rename to `skills/content/` in placement PR2 (`pzwb`)
7	core’s `catalogue`, `dublin-core`, `fhir-artifact-index` schemas	split: the generic shape stays in core; IRIS/FHIR specifics go to who-iris / fhir-harness and register on load

Other audit findings, now on the checklist:

cat-harness-tools.json needs → ["cat-harness","bootstrap-tools"], and core adds cat-harness-tools (C1, bean rmi6). 0 of tools’ 27 files import core.
34 cat-harness skills that point at sci and are in no PR yet (mostly authoring-core, sdlc-core)
90 code files that name higher instances by literal path → resolve through the checkout overlay (PR0a)

S0 rulings (2026-10-01)

Submodule pins: restore PR0’s bootstrap 7a91356 / bootstrap-tools c5e5e25. All of main’s artefacts were generated at those pins; a later merge had regressed the gitlinks.
READMEs: each repo owns its own. folio-assistant stops regenerating and checking submodule READMEs. Each repo’s own CI keeps its README current (D3). Nothing is pushed upstream.
W3C library sources (PROV-O, ODRL 2.2, JSON-LD 1.1): methodologies cite them via evidence:, reusing sibling PR #1769’s methodology nodes. The test rule stays as it is.

Rulings 2026-10-01 late

Owner, about 17:30. Each ruling is recorded in full on the bean it settles.

question	ruling	bean
Q-A generator retargeting	Split bootstrap outputs by kind. kg-qa verdicts, glossary ledgers and detangle stay in cat-harness as the auditor’s output (“a harness may discuss its dependencies”). bootstrap-tools renders the UML of bootstrap and bootstrap-tools at build time. bootstrap’s process translations move into bootstrap. Committed renderings become deploy-time output and keep their URLs. Delete the 37 duplicate sidecars. A `tools/index.ts` name collision throws. SKOS URLs (X2) are not ruled yet.	`cmsl`, `wggr`
Q-B reference direction	Q1 is reword: files stay put and their prose stops naming higher layers. X3 exempts the 3 layering specifications. The ratchet also covers single-name files (A.10). Translation mirrors are exempt (X1).	`zhg2` (implementation: `1bvx`, `vzo5`)
#1735 who-iris prose	Reword the outdated `dependents` prose in the who-iris files now, as the earlier ruling on #1735 said.	`ga6u` (`qsx4` exists only on #1735’s branch)
S5 1a, trap 1	`inProcess("src/tools/...")` resolves against the implementing instance (cat-harness-tools), found through `needs`. No cat-harness-tools path is written into cat-harness. Two points are still open: how a sidecar’s `source_file` resolves, and whether sci-bound files move in 1a.	`70lx`
merge-treadmill speed-ups	Four approved: input-hash skip (`f017`), parallel checks (`v3nf`), CI `merge:main` (`d33q` part B), and CI sharding with a BPMN cache and shallow checkout (`dlqu`).	`0mf0`

S4 evidence (#1776, the upward-links PR)

R5 upward markdown links 26 → 0 · R6 root links (md) 12 → 0 · large-datasets → core import 1 → 0 · check:reference-direction 1667 → 1640 occurrences. The 6 R6 .ts items are paths written into emitted files (init-folio.ts:279,290,318,344, translate-kg-viewer.ts:153, translation.ts:225), not root references. One upward name remains in large-datasets (gen-id-lookup.ts SOURCE = "who-iris", plus prose naming core); it goes with the needs change.

Upstream fixes (2026-10-01): done

Owner: “also fix any upstream issues”. Both PRs were merged with every job green, and both repos’ main is green on Check and Pages.

repo	PR	new `main`	what
litlfred/bootstrap-tools	#4	`30464126ed93`	`term-links.ts`: outside bootstrap, cross-repo terms link to the published IRI (`https://litlfred.github.io/bootstrap/schemas/#<term>`), never to `../bootstrap/…`, so the README is byte-identical standalone and hosted. New `check.yml` runs the README checks standalone.
litlfred/bootstrap	#1	`f75a2167d226`	new `check.yml` checks out bootstrap-tools beside it and checks the READMEs, subgraph READMEs and schema pages

Live render. The proxy blocks the live host, so each repo’s real gh-pages branch was built locally with Jekyll: index.html redirects to README.html, the generated-by notices and the no-edit footer are present, and no ../bootstrap/ links remain.
Follow-up after S0: bump folio-assistant’s pins to f75a216 / 3046412, re-extract the bootstrap translation templates, and regenerate the bootstrap kg-qa sidecars.
Minor: bootstrap-tools publishes scripts/templates/site/default.html raw, so its Liquid is visible at that URL. Exclude scripts/templates/ in site.ts.
Minor: bootstrap-tools’ unit tests need bootstrap beside them (19 fail standalone), so its CI runs only the README checks.