Implement the Harvard data sub-language: add{ ... } blocks #94

Split out of #88(B). conformance/valid/v11_add_block.tangle — def x = add{ 1 + 2 + 3 } — is the last program in the corpus the parser rejects, and it is the only one that is not a parse-rule gap.

add{ … } is a sub-language, specified across 288 lines and 12 sections of docs/spec/FORMAL-SEMANTICS.md (§6.1–§10).

What the spec requires

§6.2 — its own expression grammar (⊢_hd terms, not Tangle expressions):

he ::= n | "s" | true | false          -- literals
     | x                                -- variable (resolves in Π)
     | he₁ op he₂                       -- arithmetic, op ∈ {+,-,*,/}
     | he₁ == he₂                       -- equality
     | he₁ && he₂ | he₁ || he₂          -- boolean operators
     | !he                              -- negation
     | f(he₁, …, heₖ)                   -- call (f must be in Π)
     | if he₁ then he₂ else he₃         -- conditional, total (both branches)

Explicitly: "NO side effects, NO loops, NO assignments. Guaranteed terminating (D2.1)."

§7.1 — its own type system: Int, Float, Rational, Hex, Binary, Bool, String, Symbolic.

§7.2 — Embed / Unembed between Harvard and Tangle types (Embed(Int) = Num, Embed(Bool) = Bool, …), in both directions (§9.4).

§8 — extended environments: the Π environment (§8.2) with visibility rules (§8.1).

§9.3 — a separate typing judgement ⊢_hd, distinct from HasType.

§9.5, §9.6 — Harvard calling Tangle (D2.9) and module imports (D2.6, D2.11).

§10 — extended operational semantics.

Note harvard{ hp } (the control block, §6.3) is specified alongside it and is a further, statement-level construct.

Why it needs to be its own piece of work

None of the above is reachable by adding a grammar production. It needs new AST, new tokens (if/then/else, &&, ||, !), a second typing judgement, a conversion layer, and an evaluator — and a decision about how much of it to build.

A partial implementation is worse than none here: shipping add{ 1 + 2 + 3 } alone would claim support for a construct whose spec promises conditionals, a distinct numeric tower, and cross-language calls. The corpus file would go green while the feature stayed absent — the exact fake-gate pattern #89 was built to prevent.

Suggested shape

1. Decide the scope first. Is the Harvard layer still wanted as specified? It predates the KRL de-conflation (docs: de-conflate Tangle from KRL; drop two dead Jekyll Pages workflows #78), and ECHO-TANGLEIR-THREADING.md is from the same era — worth confirming it is not similarly stale before building 288 lines of spec.
2. If yes: the arithmetic/boolean fragment + Embed is a coherent first rung, with if/then/else, Π visibility, and calling as later ones.
3. If no: correct the spec and remove v11_add_block from the conformance corpus rather than leaving a permanently-failing "valid" program.

Until then it is the single entry in scripts/check-corpus.sh's CONFORMANCE_KNOWN_UNPARSED, so it cannot be forgotten and cannot regress silently.