Cross-project dependencies — toesnail × `.mw` × collAIb

How the three projects’ progress depends on each other, and how strongly. This is the tracking artifact for the parked collAIb × toesnail × .mw scoping session (TODO id:921b). Canonical copy lives here in toesnail (the hub of the relationship); review and update it at that session.

The three nodes

toesnail (this repo) — the theoretical-physics content / research. The use-case.
.mw (mathematical-writing) — a future literate format/VS-Code tool keeping prose + computation + machine-checked proofs mutually consistent.
collAIb (~/src/collaib) — a PWA where a local LLM watches the document and offers brief, non-intrusive side-panel observations (“calm co-author over your shoulder”).

Dependency edges

A → B = “A’s progress needs / leans on B”. Strength: none / weak / medium / strong. Blocking? = does B being absent stop A today?

From → To	Strength	Blocking?	Nature of the dependency
toesnail → `.mw`	weak	no	By design (decision D2): toesnail is authored now in plain markdown with `verify:` markers; `.mw` would make the verification workflow ergonomic, but toesnail progresses fully without it. Deliberately decoupled.
`.mw` → toesnail	strong	no (it’s a requirements dep, not a build dep)	toesnail is `.mw`’s north-star / acceptance corpus: it drives what `.mw` must support (tiered SymPy/Lean verification of real derivations, the handle model, staleness across prose↔proof). `.mw`’s design is validated against toesnail content.
toesnail → collAIb	weak	no	collAIb could provide a live in-editor UI surfacing `verify:` rigor-debt while authoring — a nice-to-have. toesnail works with `grep -rn 'verify:'` today.
collAIb → toesnail	none	n/a	collAIb is a general writing tool; toesnail is just one possible document it could observe.
`.mw` ↔ collAIb (content layer)	weak — RESOLVED (siblings)	no	Resolved at the scoping session (2026-06-15, `id:921b`). They do not integrate directly; both are independent consumers of toesnail’s plain-text marker/handle/tier schema (toesnail owns it, in `CONVENTIONS.md`). Share data, not UI — a shared UI component across the PWA/VS-Code platform line was rejected.
`.mw` ↔ collAIb (runtime layer)	PARKED — not decided	no	Three couplings flagged as questions to the owners, not decisions toesnail imposes (see “Parked runtime questions” below). Only one coherent direction emerged: `.mw` owns a headless protocol-driven verification core; collAIb’s PWA could be an alternate selectable front-end. Parked ≠ chosen — none is committed; revisit on the tripwire below.

Picture

        requirements / north-star (STRONG)
   .mw  ◀───────────────────────────────────  toesnail
    │                                            ▲
    │  overlapping authoring-surface             │  live verify: assist UI
    │  (MEDIUM, UNRESOLVED) ◀───┐                │  (WEAK, optional)
    ▼                          │                │
 collAIb  ─────────────────────┘────────────────┘

toesnail is the source of requirements but is operationally independent of both tools (weak outgoing edges) — it can keep progressing in plain markdown regardless of either tool’s state.
.mw and collAIb are tools that serve the toesnail use-case.
The .mw ↔ collAIb edge is resolved at the content layer (siblings) and parked at the runtime layer (see below) — settled at the 2026-06-15 scoping session (docs/meeting-notes/2026-06-15-1351-mw-collaib-frontend-scoping.md).

Parked runtime questions (routed to the tool owners — questions, not decisions)

The scoping session found only one coherent integration direction and parked it as three orthogonal questions, each routed to the owner repo (toesnail does not decide these):

collAIb PWA as an alternate front-end to .mw. Precondition .mw should want anyway: a front-end-agnostic, protocol-driven verification core (generalize the persistent-Lean-server dispatch into a front-end⇄core protocol). Tension: .mw explicitly names “web-hosted runtime” a non-goal — relaxing it is .mw’s owner’s call.
collAIb’s LLM observer as an optional tier-0 plugin inside .mw — advisory verify:-marker suggestions only, never a machine-checked badge (so it can’t pollute the SymPy/CrossHair/Lean tiers).
Sibling contract (= the resolved content-layer edge): always true regardless of 1 & 2.

Tripwire to revisit 1 & 2: when liveness becomes a real need (collAIb flagging a marker stale on edit, pre-save) OR when .mw ships a headless front-end⇄core protocol seam.

Witnessed instances (real evidence for the parked scoping, `id:921b`)

2026-06-15 — /relay human Resogram triage exercised the .mw → toesnail STRONG edge live. Owner ratified three SymPy-located discrepancies; applying them produced exactly the two things .mw/collAIb exist to automate:
1. Prose↔computation staleness. Adopting the exact cval form changed a claim and orphaned the surrounding prose still citing the now-removed constant c — a human had to hand-detect it and open a REVIEW_ME box. This is verbatim .mw’s id:dae5 differentiator (shipped the same day): editing a definition marks the claim AND the prose citation that desugars from it STALE, unrelated prose untouched. The cval case is a ready-made acceptance witness for that DAG — now realized in .mw as examples/resogram_cval.mw + tests/test_resogram_cval_staleness.py (.mw id:b7b1): the differentiator passes on this real content, and an xfail(strict) RED-specs a gap the episode exposed — stale_after_edit catches a changed definition ( $e$ citation + downstream average go stale) but NOT a dangling citation (the removed constant c orphans its $c$ prose citation). Authoring that witness also surfaced a second, unrelated .mw fix (id:d80d): the eval/DAG namespace treated every SymPy-callable name as a function, so the physics variable beta resolved to SymPy’s Beta function (int * FunctionClass). Fixed by a call-position rule (a name is a function only where applied, name(…)), so beta/gamma/zeta are usable symbols. → real dogfooding moves the tool, not just validates it.
2. verify→verified re-pin cascade. Correcting the edot/cval source left the verify instruments + pinned test_verify.sh verdicts + attestations stale; re-deriving them was deferred to “next review” by hand. .mw’s eval-core + staleness would re-run the verification fragment and flip the badge automatically (toesnail’s own id:04bb staleness-checker is the plain-markdown stopgap for this).
3. collAIb tripwire HIT. The parked runtime Q2 (collAIb LLM observer as an optional tier-0 advisory plugin) and the “live verify: assist UI” edge both describe precisely what would have surfaced (1) at edit time, pre-save — “this paragraph still references c, which your edit removed.” This is a concrete instance of the revisit-tripwire (“liveness becomes a real need”).

Current tool-capability map & the two-tier detection proposal (2026-06-15)

Grounded snapshot of what each sibling can do for toesnail today (read from the repos 2026-06-15, after the Resogram energy-chain episode), and the proposal that fell out of it. Meeting task queued for next session — see TODO id:d8bf. This is design analysis, not a committed plan.

`.mw` today (v0.3.0) — the HARD / deterministic tier

Usable now, headless library: parser + SymPy evaluator + content-addressed cache + a dependency-DAG with 3-state staleness (PENDING/PROVEN/STALE). stale_after_edit(old, new) returns the stale fragment set.
Proven on this content: examples/resogram_cval.mw + tests/test_resogram_cval_staleness.py pass — the differentiator works on the real Resogram episode (edit esol → its inline $e$ citation + the downstream ē average flag stale; unrelated y untouched).
Caveats: (1) No Markdown importer → can’t lift physics/Resogram.md losslessly; using it means a hand-maintained .mw mirror (double-entry — the D2 decoupling tax; cheap for one section, not repo-wide). (2) No preview / no Lean / no dispatcher (all deferred HARD). (3) Dangling-symbol gap xfail id:b7b1: it keys on changed definitions, not removed symbols — so it would NOT have caught the c-removal orphan (the origin of this episode). Owner design call pending in .mw, blocking a merge there.

collAIb today (v0.2.0) — the SOFT / advisory tier

Runnable PWA, observer loop live: local-LLM (Ollama/OpenAI-compatible) streams brief observations on a 2s-idle debounce. Works now.
But: watches only its in-browser Tiptap buffer — no filesystem access, no verify-marker awareness, no diff-aware “you removed c” check wired. So it cannot watch Resogram.md today; the only current path is copy-paste a section for generic prose feedback (low value for physics-specific needs).
The 80%-built valuable use: change-tracker.js (detects deleted spans) + observation.js (typed, confidence-thresholded, exact-substring-anchored observations) already exist and are tested but unwired. The tripwire use-case (“flag the dangling-c at edit time, pre-save”) needs file-I/O + a prompt tweak + wiring those two — a ~2–3 day feature build (the parked runtime Q; tripwire HIT this episode, see above).

The synthesis: two complementary detection tiers on the same commit-diff

The episode’s manual hunt split cleanly into what each sibling automates:

	HARD tier (deterministic)	SOFT tier (advisory)
tool	`.mw` DAG (`stale_after_edit`)	collAIb’s observer brain (prompt + local LLM)
catches	changed-definition staleness (`esol`→`ē`)	dangling refs (`c`), exposition, “asserted-not-shown”
usable today as	headless library (needs a `.mw` mirror)	headless script (needs prompt + Ollama)
output	machine-checked stale-set	advisory notes (never a badge)

Proposal — a relay-aware commit-hook (NOT yet decided; meeting topic id:d8bf)

Both tiers fire on the commit diff, emit findings into REVIEW_ME / inline 🚧 callouts, and the owner responds with the **re** (status:) vocabulary — the full detect → respond loop, automating most of this session’s manual hunt. Key realization: you don’t need either sibling’s heavy shell — not .mw’s importer nor collAIb’s PWA file-I/O. You need .mw’s DAG as a library (= toesnail’s deferred id:04bb staleness checker) and collAIb’s brain as a headless diff-observer script. toesnail’s git-commit workflow (not live-typing) makes the commit-hook the right trigger, not an editor.

“Relay-aware” = the hook must NOT fire/interfere during relay executor or pool worktree commits (they commit in worktrees outside the repo tree; the hook would either double-run or fight the integrator). Gate it on: not inside a relay worktree, no live lease held, or an explicit skip env flag. Open design questions for the meeting: pre-commit vs post-commit vs post-merge; the .mw-mirror double-entry cost vs waiting for an importer; soft-tier noise control / confidence threshold; and whether id:b7b1 (dangling-symbol detection) should be routed to this soft tier rather than .mw’s deterministic core (this episode argues yes). Weigh against the observe-before-preventing heuristic — this episode is N=1; the hook is partly its own logger.

What to watch

Keep toesnail’s outgoing edges weak on purpose — if toesnail ever requires .mw to make progress, the decoupling decision (prior D2) has been violated; revisit deliberately, don’t drift into it. The commit-hook proposal (id:d8bf) must preserve this: .mw/collAIb stay optional assists, never a gate.
If this map proves useful, mirror a one-line pointer to it from .mw’s and collAIb’s CLAUDE.md (not yet done — observe-first; do it when the relationship is confirmed stable).

2026-07-02 Fable consulting cross-dive: a full CODE/TODO/PROSE classification of every triad + satellite edge (chidiai, zkm, collAIb spikes, leAIrn2learn, Diplomarbeit), a meeting-order plan, and a proposal (P6) to extend THIS map + resolve the hub ambiguity lives at mathematical-writing/docs/meeting-notes/2026-07-02-1000-fable-consulting-grand-truth-triad.md. It also found one live twin-drift instance (collaib id:f68b’s gate id:47d4 shipped 2026-06-25, unannotated).

Cross-project dependencies — toesnail × .mw × collAIb