docs: ADR-0016 cross-language reference resolver + stack-graphs in landscape

[jamestexas]

Summary

Sets the formal model for cross-language references in mache (the work tracked under bead mache-q43l). Drafted by a math-friend / paradigm-assessor / code-architect subagent review of the existing one-arm `mod:` resolver and the original bead framing.

What landed

docs/adr/0016-cross-language-reference-resolver.md — new ADR. Three corrections vs. the original proposal:

1. Three orthogonal axes — naming, resolution, fidelity. The original collapsed them into one `scheme:value` string; separating gives independent extension axes without combinatorial blowup.
2. Resolution is a Kleisli arrow into an effects monad `T = Maybe ∘ Fetching ∘ Forbidden ∘ StaleAt(σ)`, not `Address → Graph`. The pure signature couldn't express in-flight fetches, capability denials, or arena-clock staleness — the single-arm resolver was hiding these.
3. Categorical model is a Grothendieck fibration `p: GraphMounts → Σ`, not a sheaf. Sheaf gluing requires strict overlap equality which doesn't hold under lazy fetches; honest framing is separated presheaf on a coverage, sheafified by ADR-0014's `current_root` advance. Today's `CompositeGraph` prefix-mount is the colimit of fibers — a coproduct — and the fibration adds the morphisms between fibers, which is what cross-scheme refs actually are.

Fidelity extension — cross-language refs carry fidelity at both endpoints; product poset `L_src × L_tgt` extends ADR-0013's chain mechanically. `v_defs` / `v_refs` schemas gain `target_lang` and split `fidelity` into src/tgt components. L₂ saturates at L₁ across language boundaries until a cross-language SSA / runtime-trace producer exists.

Pitch corrections (in the ADR's § Decision — Pitch and scope honesty):

• Drop "we project, we don't index" from the public pitch. Real at the algebra/coalgebra layer; identical at the runtime-cost layer. Keep as a technical remark only.
• "Monorepo = polyrepo with different URI schemes" is true at the naming layer only. False at the resolution / refactor / atomicity layers. Pitch unqualified obscures the eventual-consistency engineering the fibration model exists to make honest.

Implementation strategy — 5-bead sequence already filed under epic `mache-q43l`:

```
mache-bd97d9 (internal/resolve/ — Resolver iface + Registry)
├── mache-bdcd2b (LocalPathResolver)
│ └── mache-be0b9f (wire resolve_ref → mount in CompositeGraph) ← Terraform E2E milestone
│ └── mache-be3da9 (callers/ virtual entries)
└── mache-be650a (second scheme — npm or git remote-hint)
```

V1 keeps the resolver codomain as `(graph.Graph, error)`. The full effects monad surface lands when the first consumer needs more than `Now` / `Missing` — additive, not breaking.

`docs/competitive-landscape-2026.md` — added section 10. stack-graphs (GitHub). Both the math-friend (closest published categorical model for intra-language scope composition) and the paradigm-assessor (closest unmentioned competitor — declarative scope rules, polyglot composition, graph-composable) independently flagged stack-graphs as the most important omission. Deliberately not added to the 11-column matrix; the matrix categories don't capture what makes stack-graphs different (formal scope-graph model, intra-source polyglot). Pre-cites ADR-0016 so readers can follow the positioning argument.

`docs/ARCHITECTURE.md` — ADR table de-duplicated (the renumbered 0015 entry was sitting in the old 0006 slot AND in numerical order after 0014; collapsed to one) and ADR-0016 row added.

Subagent reports (for archival)

The fibration / product-poset / Kleisli-arrow framing came from a theoretical-foundations-analyst review of the original bead. The scope-honesty corrections came from a paradigm-assessor red-team of the "polyglot reference resolver" pitch. The 5-bead sequence came from a code-architect review that read the actual stub in `cmd/serve_resolve_ref.go` and proposed the minimal extension path. Synthesis comment lives on bead `mache-q43l`.

Test plan

• `task docs:lint` passes (0 fail, 0 warn)
• ADR-0016 cross-references resolve (ADR-0013, ADR-0014, ADR-0015 links work)
• competitive-landscape section 10 renders cleanly in GitHub
• ADR table in ARCHITECTURE.md has no duplicate 0015 row