{"row":{"id":"ADR-037","title":"Bottom-up diamond progression (no feature-milestone defer)","file":"adr-037-bottom-up-diamond-progression.md","status":"accepted","date":"2026-04-17","cites":6,"supersedes":[],"superseded_by":[],"layers":["L0","L0.5","L1","L2","L3","L4","L5"],"tags":["philosophy","roadmap","admission","forever-v0x"],"room":true},"pin":{"release_tag":"v0.107.0","commit":"5d61b7e13f1f6a10c17414f61146648f76bcf2bd"}}

[{"k":"h","depth":1,"text":"ADR-037: Bottom-up diamond progression","id":"adr-037-bottom-up-diamond-progression"},{"k":"quote","inline":[{"k":"strong","text":"⚠️ AMENDED 2026-06-20 (D-2026-06-20-N1) · the VERSION-policy layer of this ADR is superseded."},{"k":"text","text":"\nThe "},{"k":"strong","text":"core decision is INTACT"},{"k":"text","text":" · bottom-up layer progression · no\nfeature-milestone defer · every feature built as its layer is reached. What\nchanges · the \"forever-v0.x · no v1.0 target ever\" framing (ADR-002) is RETIRED\n→ real semver toward a "},{"k":"strong","text":"1.0"},{"k":"text","text":" launch. The "},{"k":"code","text":"v0.8X.Y"},{"k":"text","text":" layer-tag scheme below is a\nhistorical record · the engine is now "},{"k":"strong","text":"0.90.0"},{"k":"text","text":" (RC-grade) → "},{"k":"code","text":"1.0.0-rc.N"},{"k":"text","text":" →\n"},{"k":"strong","text":"1.0.0"},{"k":"text","text":" launch → 1.x minors add the remaining crates additively toward the\n42-crate target → "},{"k":"strong","text":"2.0"},{"k":"text","text":" = the Connectome era. The bottom-up "},{"k":"em","text":"ordering"},{"k":"text","text":" still\ngoverns which crate admits next; only the public version number now CLIMBS to\nsignal maturity instead of sitting at 0.x forever. Original body preserved per\ncross-source-validation §2.7 (zero silent rewrite)."}]},{"k":"h","depth":2,"text":"Status","id":"status"},{"k":"p","inline":[{"k":"strong","text":"Accepted"},{"k":"text","text":" 2026-04-17 night post-brainstorm (Q-plan 1a/2a/3b/4b/5-both/6a). Amends ADR-028. Cascade: ROADMAP restructure, FCI-007 re-eval, crate count target revised 40-42 → 50-90 (cap 100 unchanged)."}]},{"k":"h","depth":2,"text":"Context","id":"context"},{"k":"p","inline":[{"k":"text","text":"Nika Diamond shipped 6 admitted L0/L0.5 crates + 1 WIP ("},{"k":"code","text":"nika-schema"},{"k":"text","text":") by 2026-04-17. Throughout the first 6 admissions, the ROADMAP carried arbitrary version gates:"}]},{"k":"list","ordered":false,"items":[[{"k":"code","text":"v0.81"},{"k":"text","text":" — Foundation"}],[{"k":"code","text":"v0.90"},{"k":"text","text":" — Ship (pck + providers)"}],[{"k":"code","text":"v0.95"},{"k":"text","text":" — Cortex (9 memory satellites)"}],[{"k":"code","text":"v0.100"},{"k":"text","text":" — Plugins (WASM host + sandbox)"}]]},{"k":"p","inline":[{"k":"text","text":"These gates encoded a "},{"k":"strong","text":"feature-milestone scheduling"},{"k":"text","text":" model: ship core, then Cortex later, then plugins later. In practice this produced:"}]},{"k":"list","ordered":true,"items":[[{"k":"strong","text":"Stubs"},{"k":"text","text":" that pretend to implement reserved traits ("},{"k":"code","text":"MemoryStore"},{"k":"text","text":" in "},{"k":"code","text":"nika-kernel"},{"k":"text","text":" with no satellite impl in sight until v0.95)."}],[{"k":"strong","text":"Version-labelled drift"},{"k":"text","text":" — docs + specs referenced \"target design — arriving at v0.95\" as if it were a commitment."}],[{"k":"strong","text":"Shadow zones"},{"k":"text","text":" — the gap between \"reserved\" and \"implemented\" accumulated residue (invariants drift, type changes needed later)."}],[{"k":"strong","text":"Fake ambition"},{"k":"text","text":" — any mention of \"we'll ship X at vN.M\" is a promise we cannot keep with forever-v0.x."}]]},{"k":"p","inline":[{"k":"text","text":"User said 2026-04-17 night:"}]},{"k":"quote","inline":[{"k":"text","text":"\"c'est une mauvaise chose de le faire en différé. on a pas la bonne logique v0.95 v0.100. je veux vraiment qu'on fasse toutes les features au fur et à mesure en partant du bas du diamant et on remonte.\""}]},{"k":"p","inline":[{"k":"text","text":"Translation: "},{"k":"strong","text":"defer = bad logic. Build every feature as its layer is reached, from the bottom of the diamond upward."}]},{"k":"h","depth":2,"text":"Decision","id":"decision"},{"k":"p","inline":[{"k":"text","text":"Abandon feature-milestone scheduling. Adopt "},{"k":"strong","text":"bottom-up layer progression"},{"k":"text","text":":"}]},{"k":"code","lang":"text","text":"L0    (~9 crates)     foundation types, pure, sync\n  ↓\nL0.5  (~2 crates)     kernel traits (sealed, ISP, async OK in defs)\n  ↓\nL1    (~20+ crates)   effect impls — fs, http, process, git, memory-*, provider-*\n  ↓\nL2    (~10+ crates)   verbs + domain services + memory orchestrator + pck\n  ↓\nL3    (~5 crates)     runtime + shield + wasm-host + sandbox\n  ↓\nL4    (~7 crates)     interfaces — cli, lsp, daemon, mcp, sdk, serve\n  ↓\nL5    (1 crate)       the `nika` binary"},{"k":"p","inline":[{"k":"text","text":"Every crate admits atomically with 12 gates. Every admission = 1 commit = 1 "},{"k":"code","text":"v0.80.N"},{"k":"text","text":" tag increment."}]},{"k":"h","depth":3,"text":"What changes","id":"what-changes"},{"k":"list","ordered":true,"items":[[{"k":"strong","text":"ROADMAP.md restructure"},{"k":"text","text":" — replace "},{"k":"code","text":"v0.81/v0.90/v0.95/v0.100"},{"k":"text","text":" section headers with layer-progression sections. No calendar targets. No \"this ships at vN.M\" language."}],[{"k":"strong","text":"ADR-028 re-scope"},{"k":"text","text":" — distinguish:"},{"k":"text","text":"- **API type reservation** (non_exhaustive ratchet, reserved fields in InferRequest/InferResponse, kernel trait surface) — KEEP. Legitimate forward-compat work.\n- **Feature scheduling** (we'll ship Cortex at v0.95) — REMOVE. Bad diamond."}],[{"k":"strong","text":"FCI-007 `unstable-cortex` re-evaluate"},{"k":"text","text":" — if Cortex ships layer-by-layer without defer, we don't need a feature flag gating it. Either drop or repurpose as per-crate experimental opt-in."}],[{"k":"strong","text":"Version tagging — Q-plan 3b LOCKED"},{"k":"text","text":": "},{"k":"code","text":"v0.8X.Y"},{"k":"text","text":" where X climbs per layer-phase completion, Y increments per admission within a phase."},{"k":"text","text":"- `v0.80.N` → current baseline (L0 partial, L0.5 complete)\n- `v0.81` → L0 complete (9 crates admitted)\n- `v0.82` → L1a primitives complete (fs, process, http-client, keys-env)\n- `v0.83` → L1b deps-étage-1 (git, keys-keychain, catalog-sync)\n- `v0.84` → L1c providers (rig + openai-compat + mock)\n- `v0.85` → L1d pck backend (registry + store)\n- `v0.86` → L1e memory foundation (embed, store-fjall, autodesc, hnsw, bm25, rrf)\n- `v0.87` → L1f memory advanced (temporal, fsrs, rdfs-reasoner, graph-algos)\n- `v0.88` → L1g reserved satellites (as demand emerges: causal, episodic, working, meta, procedural, spatial)\n- `v0.89` → L2 complete (verbs, orchestrators, builtins)\n- `v0.90` → L3 complete (runtime, shield, wasm-host, sandbox). **NOT a \"ship\" tag** per Q-plan 6a.\n- `v0.91` → L4 complete (cli, daemon, serve, mcp, lsp, sdk)\n- `v0.92` → L5 binary first full-stack integration\n- `v0.93+` → polish, optimization, quality iterations\n- ~~Forever-v0.x — no v1.0 target ever.~~ (amended D-2026-06-20-N1 · real semver toward a 1.0 launch · this layer-tag scheme is now a historical record · see top banner)"}],[{"k":"strong","text":"Total crate count target"},{"k":"text","text":" — Q-plan 1a LOCKED: 40-42 → "},{"k":"strong","text":"50-90 crates"},{"k":"text","text":" (cap 100 unchanged). Growth driven by 14-crate memory subsystem, 4 storage backends, ~30 provider crates, own "},{"k":"code","text":"nika-embed"},{"k":"text","text":", WASM host + 3 sandbox crates."}],[{"k":"strong","text":"Dedicated deep-dive sessions reserved"},{"k":"text","text":" — Q-plan 5 LOCKED both:"},{"k":"text","text":"- **Session nika-memory** — most-ambitious B1-B5 validated 2026-04-17 night. Design 14 crates (8 core + 6 reserved). Runs FIRST (stabilizes kernel memory traits before L1 memory admission).\n- **Session nika-registry** — backend + publishing policy. Runs after memory."}],[{"k":"strong","text":"\"v0.90 ship target\" DROPPED"},{"k":"text","text":" — Q-plan 6a. Versions 0.9X are just layer tags, no marketing ceremony. (Amended D-2026-06-20-N1 · the "},{"k":"strong","text":"1.0.0"},{"k":"text","text":" launch is now the marketing moment · 0.9X stays RC-grade pre-launch · per top banner.)"}]]},{"k":"h","depth":3,"text":"What stays the same","id":"what-stays-the-same"},{"k":"list","ordered":false,"items":[[{"k":"text","text":"12-gate admission model (unchanged)."}],[{"k":"text","text":"28+ invariants (all still binding)."}],[{"k":"text","text":"7 shadow zones (still pre-launch quality gates)."}],[{"k":"code","text":"publish = false"},{"k":"text","text":" on foundation crates (ADR-022 unchanged)."}],[{"k":"text","text":"The release model (ADR-002) — amended D-2026-06-20-N1 to real semver toward a 1.0 launch (was \"forever-v0.x · no v1.0 target\")."}]]},{"k":"h","depth":3,"text":"Ordering within a layer","id":"ordering-within-a-layer"},{"k":"p","inline":[{"k":"text","text":"Dependencies-first topological order. If two crates have no dep relation, tiebreaker:"}]},{"k":"list","ordered":true,"items":[[{"k":"strong","text":"User-value-first"},{"k":"text","text":" (which unlocks a demo or blocks downstream first)."}],[{"k":"strong","text":"LOC-smallest-first"},{"k":"text","text":" (easier wins early build momentum)."}],[{"k":"strong","text":"Effect-axis-smallest-first"},{"k":"text","text":" (fewer capability axes = lower review burden)."}]]},{"k":"p","inline":[{"k":"text","text":"Example L1 candidates in topological order:"}]},{"k":"list","ordered":false,"items":[[{"k":"code","text":"nika-fs"},{"k":"text","text":" (L1, reads-fs rw-fs, no deps on other L1)"}],[{"k":"code","text":"nika-process"},{"k":"text","text":" (L1, exec-shell, no deps)"}],[{"k":"code","text":"nika-http-client"},{"k":"text","text":" (L1, net-egress, no deps)"}],[{"k":"code","text":"nika-git"},{"k":"text","text":" (L1, depends on "},{"k":"code","text":"nika-fs"},{"k":"text","text":" + "},{"k":"code","text":"nika-process"},{"k":"text","text":")"}],[{"k":"code","text":"nika-memory-*"},{"k":"text","text":" (L1, each independent, 8 core + reserved)"}],[{"k":"code","text":"nika-<provider>-*"},{"k":"text","text":" (L1, depend on "},{"k":"code","text":"nika-http-client"},{"k":"text","text":")"}]]},{"k":"h","depth":2,"text":"Consequences","id":"consequences"},{"k":"list","ordered":false,"items":[[{"k":"text","text":"✅ Bottom-up integrity — no fake feature-milestones, no version-labelled residue (the anti-defer discipline survives the D-2026-06-20-N1 real-semver amendment; only the version labels changed, not the build order)."}],[{"k":"text","text":"✅ Every admitted crate is REAL code, not stub. No \"v0.95 target design\" banners in docs."}],[{"k":"text","text":"✅ Linear ROADMAP — users see \"what's admitted, what's coming up in this layer, what's above\"."}],[{"k":"text","text":"✅ Tagging discipline — "},{"k":"code","text":"v0.80.N"},{"k":"text","text":" monotonic, no backtracking, no retroactive milestone rewrites."}],[{"k":"text","text":"❌ Harder to communicate \"when does Cortex ship\" — answer is \"when we reach L1/L2 memory satellites, and never before then\"."}],[{"k":"text","text":"❌ Existing docs (ROADMAP sections, spec drafts, adr-036 MSRV body) that reference "},{"k":"code","text":"v0.90"},{"k":"text","text":"/"},{"k":"code","text":"v0.95"},{"k":"text","text":"/"},{"k":"code","text":"v0.100"},{"k":"text","text":" need cascade edits."}],[{"k":"text","text":"⚠️ Some speculative scheduling (nika-sdk crates.io publishing, MSRV cadence) currently pinned to v0.90 — adr-036 must be reworded to say \"at whichever tag crosses L4 interface completion\" rather than \"v0.90\"."}]]},{"k":"h","depth":2,"text":"Alternatives considered","id":"alternatives-considered"},{"k":"list","ordered":true,"items":[[{"k":"strong","text":"Keep version-milestone gates but clarify they're goals not deadlines"},{"k":"text","text":" — fails because docs still communicate false scheduling."}],[{"k":"strong","text":"Hybrid: layer-progression as primary, version labels as secondary guidance"},{"k":"text","text":" — fails because version labels re-encode the deferred-scheduling mindset."}],[{"k":"strong","text":"Skip ADR-028 revision, just restructure ROADMAP"},{"k":"text","text":" — fails because ADR-028 still uses \"reservation for v0.95 Cortex\" language that contradicts this ADR."}]]},{"k":"h","depth":2,"text":"Cascade plan (post-acceptance)","id":"cascade-plan-post-acceptance"},{"k":"list","ordered":true,"items":[[{"k":"text","text":"Restructure ROADMAP.md (1 commit)."}],[{"k":"text","text":"Re-scope ADR-028 (1 commit, Under Revision → Amended)."}],[{"k":"text","text":"Re-evaluate FCI-007 (1 commit in "},{"k":"code","text":"forward-compat-invariants.md"},{"k":"text","text":")."}],[{"k":"text","text":"Update CLAUDE.md narrative (1 commit)."}],[{"k":"text","text":"Reserve dedicated sessions for nika-memory + nika-registry (memory files, no commits)."}],[{"k":"text","text":"Rework adr-036 MSRV body to remove v0.90 references (1 commit)."}],[{"k":"text","text":"Constellation reconciliation report: Note Option A/B/C oxigraph rename deferred to nika-memory session (1 commit)."}]]},{"k":"h","depth":2,"text":"See also","id":"see-also"},{"k":"list","ordered":false,"items":[[{"k":"text","text":"ADR-001 — Diamond orphan branch."}],[{"k":"text","text":"ADR-002 — release model (amended D-2026-06-20-N1 · real semver toward 1.0 · was \"forever-v0.x\")."}],[{"k":"text","text":"ADR-003 — 12-gate admission."}],[{"k":"text","text":"ADR-004 — Context-window-sized crates."}],[{"k":"text","text":"ADR-028 — Forward-compat reservation policy (amended by this ADR)."}],[{"k":"code","text":"philosophy_bottom_up_diamond_2026_04_17.md"},{"k":"text","text":" in memory — full brainstorm record."}]]},{"k":"p","inline":[{"k":"text","text":"🦋"}]}]

ADR-037

Bottom-up diamond progression (no feature-milestone defer)

accepted · 2026-04-17 · L0 L0.5 L1 L2 L3 L4 L5 · cites 6

ADR-037: Bottom-up diamond progression #

⚠️ AMENDED 2026-06-20 (D-2026-06-20-N1) · the VERSION-policy layer of this ADR is superseded. The core decision is INTACT · bottom-up layer progression · no feature-milestone defer · every feature built as its layer is reached. What changes · the "forever-v0.x · no v1.0 target ever" framing (ADR-002) is RETIRED → real semver toward a 1.0 launch. The v0.8X.Y layer-tag scheme below is a historical record · the engine is now 0.90.0 (RC-grade) → 1.0.0-rc.N → 1.0.0 launch → 1.x minors add the remaining crates additively toward the 42-crate target → 2.0 = the Connectome era. The bottom-up ordering still governs which crate admits next; only the public version number now CLIMBS to signal maturity instead of sitting at 0.x forever. Original body preserved per cross-source-validation §2.7 (zero silent rewrite).

Status #

Accepted 2026-04-17 night post-brainstorm (Q-plan 1a/2a/3b/4b/5-both/6a). Amends ADR-028. Cascade: ROADMAP restructure, FCI-007 re-eval, crate count target revised 40-42 → 50-90 (cap 100 unchanged).

Context #

Nika Diamond shipped 6 admitted L0/L0.5 crates + 1 WIP (nika-schema) by 2026-04-17. Throughout the first 6 admissions, the ROADMAP carried arbitrary version gates:

v0.81 — Foundation
v0.90 — Ship (pck + providers)
v0.95 — Cortex (9 memory satellites)
v0.100 — Plugins (WASM host + sandbox)

These gates encoded a feature-milestone scheduling model: ship core, then Cortex later, then plugins later. In practice this produced:

Stubs that pretend to implement reserved traits (MemoryStore in nika-kernel with no satellite impl in sight until v0.95).
Version-labelled drift — docs + specs referenced "target design — arriving at v0.95" as if it were a commitment.
Shadow zones — the gap between "reserved" and "implemented" accumulated residue (invariants drift, type changes needed later).
Fake ambition — any mention of "we'll ship X at vN.M" is a promise we cannot keep with forever-v0.x.

User said 2026-04-17 night:

"c'est une mauvaise chose de le faire en différé. on a pas la bonne logique v0.95 v0.100. je veux vraiment qu'on fasse toutes les features au fur et à mesure en partant du bas du diamant et on remonte."

Translation: defer = bad logic. Build every feature as its layer is reached, from the bottom of the diamond upward.

Decision #

Abandon feature-milestone scheduling. Adopt bottom-up layer progression:

text

1L0    (~9 crates)     foundation types, pure, sync2  ↓3L0.5  (~2 crates)     kernel traits (sealed, ISP, async OK in defs)4  ↓5L1    (~20+ crates)   effect impls — fs, http, process, git, memory-*, provider-*6  ↓7L2    (~10+ crates)   verbs + domain services + memory orchestrator + pck8  ↓9L3    (~5 crates)     runtime + shield + wasm-host + sandbox10  ↓11L4    (~7 crates)     interfaces — cli, lsp, daemon, mcp, sdk, serve12  ↓13L5    (1 crate)       the `nika` binary

Every crate admits atomically with 12 gates. Every admission = 1 commit = 1 v0.80.N tag increment.

What changes #

ROADMAP.md restructure — replace v0.81/v0.90/v0.95/v0.100 section headers with layer-progression sections. No calendar targets. No "this ships at vN.M" language.
ADR-028 re-scope — distinguish:- **API type reservation** (non_exhaustive ratchet, reserved fields in InferRequest/InferResponse, kernel trait surface) — KEEP. Legitimate forward-compat work. - **Feature scheduling** (we'll ship Cortex at v0.95) — REMOVE. Bad diamond.
FCI-007 `unstable-cortex` re-evaluate — if Cortex ships layer-by-layer without defer, we don't need a feature flag gating it. Either drop or repurpose as per-crate experimental opt-in.
Version tagging — Q-plan 3b LOCKED: v0.8X.Y where X climbs per layer-phase completion, Y increments per admission within a phase.- `v0.80.N` → current baseline (L0 partial, L0.5 complete) - `v0.81` → L0 complete (9 crates admitted) - `v0.82` → L1a primitives complete (fs, process, http-client, keys-env) - `v0.83` → L1b deps-étage-1 (git, keys-keychain, catalog-sync) - `v0.84` → L1c providers (rig + openai-compat + mock) - `v0.85` → L1d pck backend (registry + store) - `v0.86` → L1e memory foundation (embed, store-fjall, autodesc, hnsw, bm25, rrf) - `v0.87` → L1f memory advanced (temporal, fsrs, rdfs-reasoner, graph-algos) - `v0.88` → L1g reserved satellites (as demand emerges: causal, episodic, working, meta, procedural, spatial) - `v0.89` → L2 complete (verbs, orchestrators, builtins) - `v0.90` → L3 complete (runtime, shield, wasm-host, sandbox). **NOT a "ship" tag** per Q-plan 6a. - `v0.91` → L4 complete (cli, daemon, serve, mcp, lsp, sdk) - `v0.92` → L5 binary first full-stack integration - `v0.93+` → polish, optimization, quality iterations - ~~Forever-v0.x — no v1.0 target ever.~~ (amended D-2026-06-20-N1 · real semver toward a 1.0 launch · this layer-tag scheme is now a historical record · see top banner)
Total crate count target — Q-plan 1a LOCKED: 40-42 → 50-90 crates (cap 100 unchanged). Growth driven by 14-crate memory subsystem, 4 storage backends, ~30 provider crates, own nika-embed, WASM host + 3 sandbox crates.
Dedicated deep-dive sessions reserved — Q-plan 5 LOCKED both:- **Session nika-memory** — most-ambitious B1-B5 validated 2026-04-17 night. Design 14 crates (8 core + 6 reserved). Runs FIRST (stabilizes kernel memory traits before L1 memory admission). - **Session nika-registry** — backend + publishing policy. Runs after memory.
"v0.90 ship target" DROPPED — Q-plan 6a. Versions 0.9X are just layer tags, no marketing ceremony. (Amended D-2026-06-20-N1 · the 1.0.0 launch is now the marketing moment · 0.9X stays RC-grade pre-launch · per top banner.)

What stays the same #

12-gate admission model (unchanged).
28+ invariants (all still binding).
7 shadow zones (still pre-launch quality gates).
publish = false on foundation crates (ADR-022 unchanged).
The release model (ADR-002) — amended D-2026-06-20-N1 to real semver toward a 1.0 launch (was "forever-v0.x · no v1.0 target").

Ordering within a layer #

Dependencies-first topological order. If two crates have no dep relation, tiebreaker:

User-value-first (which unlocks a demo or blocks downstream first).
LOC-smallest-first (easier wins early build momentum).
Effect-axis-smallest-first (fewer capability axes = lower review burden).

Example L1 candidates in topological order:

nika-fs (L1, reads-fs rw-fs, no deps on other L1)
nika-process (L1, exec-shell, no deps)
nika-http-client (L1, net-egress, no deps)
nika-git (L1, depends on nika-fs + nika-process)
nika-memory-* (L1, each independent, 8 core + reserved)
nika-<provider>-* (L1, depend on nika-http-client)

Consequences #

✅ Bottom-up integrity — no fake feature-milestones, no version-labelled residue (the anti-defer discipline survives the D-2026-06-20-N1 real-semver amendment; only the version labels changed, not the build order).
✅ Every admitted crate is REAL code, not stub. No "v0.95 target design" banners in docs.
✅ Linear ROADMAP — users see "what's admitted, what's coming up in this layer, what's above".
✅ Tagging discipline — v0.80.N monotonic, no backtracking, no retroactive milestone rewrites.
❌ Harder to communicate "when does Cortex ship" — answer is "when we reach L1/L2 memory satellites, and never before then".
❌ Existing docs (ROADMAP sections, spec drafts, adr-036 MSRV body) that reference v0.90/v0.95/v0.100 need cascade edits.
⚠️ Some speculative scheduling (nika-sdk crates.io publishing, MSRV cadence) currently pinned to v0.90 — adr-036 must be reworded to say "at whichever tag crosses L4 interface completion" rather than "v0.90".

Alternatives considered #

Keep version-milestone gates but clarify they're goals not deadlines — fails because docs still communicate false scheduling.
Hybrid: layer-progression as primary, version labels as secondary guidance — fails because version labels re-encode the deferred-scheduling mindset.
Skip ADR-028 revision, just restructure ROADMAP — fails because ADR-028 still uses "reservation for v0.95 Cortex" language that contradicts this ADR.

Cascade plan (post-acceptance) #

Restructure ROADMAP.md (1 commit).
Re-scope ADR-028 (1 commit, Under Revision → Amended).
Re-evaluate FCI-007 (1 commit in forward-compat-invariants.md).
Update CLAUDE.md narrative (1 commit).
Reserve dedicated sessions for nika-memory + nika-registry (memory files, no commits).
Rework adr-036 MSRV body to remove v0.90 references (1 commit).
Constellation reconciliation report: Note Option A/B/C oxigraph rename deferred to nika-memory session (1 commit).