Revamps the README for the landed surface

[johnnyt]

Why

The README was written when the package was a scaffold with accepted records
and no implementation. Campaign 009 landed the core - the vault surface and
its use macro, Encryptor.Kdf including extract/2 and salted_subkey/5,
the derived-subkey vault surface, the envelope, the provider behaviour with
its conformance suite, the closed error vocabulary, the bounded cache and its
recycler, rekey - so the README's framing ("built, not released", an ADR table
as the main body) no longer describes what a reader gets.

It also gave a reader no signal about stability. The package is weeks from
1.0.0 and its storage formats and derivation constants are still moving, and
nothing in the README said so.

What

• A pre-1.0 stability notice at the top, in the operator's commissioned
  wording: APIs, storage formats and derivation constants may change between
  releases without a deprecation cycle until 1.0.0; pin an exact version and
  review the changelog before upgrading.
• A real library README structure: what it is, the security model in
  brief (the three-level hierarchy plus the six properties the package
  enforces rather than documents), installation, a quickstart, a surface map,
  derived subkeys, what does not exist yet, documentation pointers, the ADR
  table with statuses, the family split, engine notes, contributing.
• Installation now names the target: the reserved encryptor 0.1.0 on
  Hex holds no implementation, no release has been cut, and the first real
  release is 0.2.0 - so a reader knows both what to pin today (a full SHA)
  and what to move to later. This matches CHANGELOG.md, which is the
  authority on the point.
• A quickstart that runs. Every claim in it was executed against the
  compiled library rather than transcribed from memory - see Notes.
• What does not exist yet, stated rather than implied: there is no
  Argon2id surface (enc-dtv), and ADR-0006 telemetry is proposed, not
  accepted, with no events emitted. Both were previously invisible to a
  reader, and the second is the kind of gap someone builds a dashboard
  against before discovering it.
• The :derivation_salt rotation cost stated plainly: rotating it is a
  full reindex, because every value ever derived under the old salt changes.

No version bump, no mix.exs release fields touched, no ADR status changed.

Notes

Every factual claim was machine-checked against the compiled library, not
against the guides:

• The quickstart was executed end to end via mix run. The describe/1
  output printed in the README (committed?: true, encrypted_data_keys: [%{key_name: "card/v1", provider_id: "acme_payments"}], the three-pair
  context) is the actual return, and both printed refusals -
  {:missing_required_context_keys, ["column"]} and :decrypt_failed on a
  substituted context - are actual returns too.
• The derived-subkey example was executed against a vault configured with a
  :derivation_salt, as was the {:missing_config, [:derivation_salt]} path.
• Every module and function named in the surface-map table was checked with
  function_exported?/3 against the compiled beams. One probe miss was mine,
  not the README's: Encryptor.Error.describe/1 is private, and the README
  makes no claim about it.
• :telemetry is confirmed absent from the loaded applications, which is what
  makes "proposed, not emitted" a checked statement.
• The refused key-material use options, the 20 * max_age recycler default,
  the two accepted algorithm suites and the ~> 1.18 Elixir requirement were
  each read off the source rather than the prose.

Noted, not touched (outside this bead's scope, flagged rather than fixed):
CLAUDE.md still carries scaffold-era text - "Nothing is implemented yet.
The repository holds the scaffold only, so almost every convention below is
inherited rather than demonstrated" - and its "Read before writing any code
here" section still points at the ADR beads rather than the accepted
records. Same residue enc-nsi's landing flagged. It needs its own bead.

The README quickstart is transcribed from getting-started Part 1, whose code
blocks Encryptor.GuidesTest already executes; the independent mix run
above is the direct evidence for the README's own copy.

Gate: docs-only, so the kit reports no gate applicable
(carve_out_reason: no changes under lib/, test/, config/, mix.exs,
mix.lock). A full mix quality was run anyway as evidence and was green:
format, compile, deps, credo, dialyzer clean, 419/419 tests, 98.2% coverage.
Doctor, gettext and sobelow are the project's standing not-applicable skips.

Closes enc-ajj