chore: remove the bindgen prototype - #386

Ruling 2026-08-12: bindgen will not work well enough; remove it whole. Git
history keeps it — the rationale this repository recorded when PR #311 deleted
dead code.

The prototype generated Starlark bindings from cobra CLI source: an AST
extractor enumerated commands and flags (144 commands and 272 flags from
docker-cli v27.4.1 in its one recorded run, 2026-01-11), and a generator turned
the schema into receiver scaffolding. Its own README documents the ceiling that
ended it: AST extraction yields syntax without semantics — no positional
arguments, no valid flag combinations, wrong invocation forms — so every
generated command needed human refinement while fifteen hand-written ones were
production ready. The successor is AI-driven authoring against vendor
documentation: lore onboard today, the licensing draft's authoring-CAG
strategically, which gets the semantics extraction structurally cannot.

The tree was already dead by every mechanical measure, discovered when audit
phase 1b-iii went to decompose extractFromFunction: the extractor, its test,
and cmd/main.go all carry //go:build ignore; the extractor imports
internal/bindgen, which does not exist; nothing imports any of it; and go mod
tidy had already dropped golang.org/x/tools because its only consumer is
ignore-tagged. The two bare nolint directives in the extractor suppressed
nothing — golangci-lint never loads ignore-tagged files.

Phase 1b-iii of issue #365 is thereby resolved by deletion rather than
decomposition; bare nolint directives across the tree drop 5 -> 3, all of them
1b-iv/1b-v scope. Two audit-statement corrections ride in the plan: the claim
that no target path was excluded from the linter was false in effect for these
two functions (a build tag excludes as thoroughly as config, and the audit
never checked tags), and the repo-wide gocognit live-finding count was 9, not
11.

One more prediction fell to a guard during execution, and is recorded in the
plan: the verification bar said go mod tidy would be diffless, but
golang.org/x/text reclassifies from direct to indirect — a prototype file was
its last direct importer. Nothing leaves the graph; go.sum is unchanged; the
tidied go.mod rides here.

The learnings are preserved on issue #215, which this closes: the extraction
results, the README's comparison-table verdict, and this commit's SHA for
retrieval of any of it.

Verified before committing: no prototype package in go list, zero bindgen
references in Go source, exactly 3 bare nolint directives, go mod tidy
converged. The suite loses one green package (prototype/bindgen/internal) that
nothing imports.
Plan: docs/plans/remove-bindgen-prototype.md.
Issues: #215 (closed by this), #365 (phase 1b-iii)