correct five stale claims in the docs and package comments

[haribo]

Why

Five statements in the docs and package comments describe a shellf that no longer exists. Each
verified against the current tree.

where	says	actually
docs/language.md:3-4	"One shipped construct has no chapter yet — as <user>"	no chapter for on, parallel, override def, or the inventory syntax either (grep '^## ' docs/language.md)
docs/language.md:28	a def translates a ShellResult "via when/tags"	when is not a keyword — zero occurrences in internal/lang
docs/design/inventory.md:16	address | yes | network endpoint	not required for local: "true" (README.md:112). local and interpreter are absent from the table although both are host fields (internal/inventory/inventory.go:17-18)
internal/inventory/inventory.go:1-3	"Connection coordinates only — no business variables. Built in Go for now; a parsed shellf-language inventory comes later"	both clauses false: Vars map[string]string at :19, and the inventory is parsed in the shellf language at internal/lang/parser.go:361
internal/std/postgres/postgres.shellf:3	"Only these two"	four defs in the file. Worse, :5 argues that a postgres.config "would also inherit the versioned-path problem (#546), unresolved" — 130 lines above the postgres.config that resolves it with pg_conftool

docs/design/inventory.md:63 does carry a > **Superseded.** note, but it covers only the
free-form-variables point. The address/local/interpreter row is untouched by it — worth
saying, because an earlier audit in #642 read that note as covering the file and it does not.

Not a defect, listed so it is not re-raised

docs/dogfood.md:99 — "A plan cannot read another host's inventory entry" — is superseded by
ADR-0054. It stays: dogfood.md is a dated report of what a deployment could not express,
and history is allowed to describe a world the code has left (ADR-0001). A one-line forward
pointer to ADR-0054 is welcome; rewriting the finding is not.

Build

Correct the five. postgres.shellf's header is the one to think about rather than patch: it
argues a decision that the file itself has since reversed, so it needs rewriting, not a count
fix.

language.md is the largest: the missing chapters are real work, and the honest first move is
to make the header say what is missing instead of understating it. Writing the chapters can be
its own issue — but a spec that claims completeness it does not have is worse than one that
lists its holes.

Validation

• grep -n 'when' docs/language.md — no occurrence presents it as a construct.
• docs/language.md's header names every shipped construct that has no chapter, checked
  against grep '^## ' docs/language.md and the plan grammar in internal/lang/parser.go.
• docs/design/inventory.md's host table lists local and interpreter, and address is not
  marked required unconditionally.
• internal/inventory/inventory.go's package comment is true of the package as it is.
• grep -c '^def ' internal/std/postgres/postgres.shellf agrees with its header.

Out of scope

Writing the missing language.md chapters. Naming the gap is this issue; filling it is
another, and pretending otherwise is how this one never closes.

[haribo]

Delivered in #671, squashed on develop, together with #661 — the two touch the same two files
for the same reason (key is documented in the README and in the host table).

All five corrected, each re-verified on develop first:

• language.md's header named one missing chapter and was short by four (on, parallel,
  override def, the inventory syntax) — it now names all five, and says it used to understate.
• language.md:28 cited a when keyword with zero occurrences in internal/lang.
• the host table said address was always required and omitted local and interpreter.
• internal/inventory's package comment was false in both clauses.
• postgres.shellf said "only these two" with four defs, and argued that a postgres.config
  would inherit an unresolved problem a hundred lines above the one that resolves it. Rewritten
  as a prediction the code overtook, rather than corrected to "four": a comment that records what
  it believed without recording what disproved it goes stale again.

Found while fixing: README.md said "Only address is required" three lines above
"(no address needed)" for a local host — the same contradiction as the design doc, in the
file reviewed twice that day.

docs/dogfood.md:99 stays as written, deliberately, with a forward pointer to ADR-0054: it is a
dated report, and history may describe a world the code has left (ADR-0001).