feat(plan): plan mode, decision records, and the glossary

docs/adr/0001-decision-records-storage.md

@@ -0,0 +1,3 @@
+# Store decision records under docs/adr/ with an adr:<slug> citation token
+
+**Context.** Every load-bearing decision except technology had three destinations, all of which rot: "Context and vision" prose the parser ignores, `[ASSUMPTION]` markers nothing forces resolving, or nowhere at all — so the *why* never travelled with the *what* into specs and runner sessions. **Decision.** Decision records live in project-scoped `docs/adr/NNNN-<slug>.md` (append-only, one flat sequence, like `docs/stack.md`), and feats cite them with a third ref token, `adr:<slug>`, tokenized and linted beside `stack:` and `[[wiki]]`. **Why.** Reusing the proven stack-ref machinery (tokenize → validate → inline into the brief) makes hand-written ADRs first-class before any skill ships, and a lintable citation is what a plain `NNNN-slug.md` file alone lacks; the rejected alternatives — a wiki page (prose, not a contract) and a plan-scoped record (invisible to sibling plans) — both recreate the drift this feature removes.

docs/adr/0002-two-tier-interview.md

@@ -0,0 +1,3 @@
+# Draft interviews in two tiers — batch scope facts, grill decisions
+
+**Context.** The prd skill's Draft used one batched multiple-choice interview ("1B, 2A, 3C") optimized for speed, which bundles load-bearing decisions in with scope facts so they pass undiscussed; pure one-at-a-time grilling (the studied grilling skill) is the opposite failure — it makes cheap scope facts as expensive as real decisions. **Decision.** Draft runs two tiers: the existing batched lettered multiple-choice for scope facts, then a sequential Decision grill (one decision per exchange, recommendation first, dependency order, ADR written inline) for every candidate that passes the triple gate. **Why.** csdd's batch already beats grilling for facts and grilling already beats batch for decisions, so splitting on the triple gate takes the strength of each; the grill applies only to the 2–5 gate-positive decisions a typical initiative carries, bounding the interview cost that killed pure grilling.

docs/adr/0003-glossary-deferred.md

@@ -0,0 +1,3 @@
+# Defer the glossary / ubiquitous language to its own phase
+
+**Context.** domain-modeling bundles a `CONTEXT.md` glossary (ubiquitous language, canonical entity aliases) with its ADR discipline, and it was tempting to ship both under the decision-records feature. **Decision.** The glossary is deferred to its own phase (`docs/plans/PLAN-glossary.md`) rather than folded into decision-records. **Why.** The glossary's real value is the canonical/alias entity table feeding the knowledge-base graph extractor, so it must be designed *with* the graph — a different seam than the plan-grammar/validate/brief seams this feature touches; conflating the two would widen this feature's blast radius for no shared machinery. This is the explicit "no" — the deferral is a decision, not an omission — and PLAN-glossary is the phase that resolves it.

docs/glossary.md

@@ -0,0 +1,52 @@
+# Glossary — csdd's ubiquitous language
+
+This is the project's **ubiquitous-language contract**: one canonical term per
+domain concept, defined as what it **is**, with an `_Avoid_` list of the synonyms
+it bans. Identifiers (feat slugs, spec directories, wiki page names) are minted
+from canonical terms; `csdd plan validate`, `csdd graph analyze`, and
+`csdd wiki lint` report an avoided term used as a whole token. Prose outside the
+`## Language` section is ignored by the parser.
+
+## Language
+
+### Planning
+
+**Plan**: The structured document above specs (`docs/plans/<slug>/`) that
+decomposes an initiative into feats; each feat becomes exactly one spec.
+_Avoid_: prd, epic
+
+**Feat**: One row of a plan's Feats table — a spec-sized chunk of work with an
+explicit objective, dependencies, and refs; it becomes exactly one spec.
+_Avoid_: feature, story, ticket
+
+**Spec**: The per-feature contract under `specs/<name>/` (requirements, design,
+tasks) that a feat is realized as.
+_Avoid_: specification-doc
+
+**Seed**: A pre-authored artifact under a plan's `seeds/<feat>/` that hands the
+eventual spec a head start.
+
+**Brief**: The deterministic context pack `csdd plan brief` assembles for one run
+step — the feat, its refs, decisions, and the step contract.
+
+### Contracts
+
+**Quality Gate**: A `- <label>: <command>` line in a plan's Quality Gates section;
+the runner executes every gate after each implementation step.
+
+**Stack Row**: One Decided-table row of `docs/stack.md` — a one-line technology
+contract a feat cites as `stack:<name>`.
+
+**ADR**: A decision record under `docs/adr/NNNN-<slug>.md` capturing a
+gate-positive decision (hard to reverse, surprising, a real trade-off); a feat
+cites it as `adr:<slug>`.
+_Avoid_: decision-doc
+
+### Knowledge base
+
+**Steering**: The project-memory documents under `.claude/steering/` (product,
+tech, structure) that source facts for the authoring flows.
+
+**Wiki Page**: One concept page under `docs/wiki/pages/<slug>.md`, derived from a
+raw source and cross-linked with `[[wikilinks]]`.
+_Avoid_: article