platform: extend or formally retire the string/boolean-only COMMAND_PROTOCOL_V1 #396

Goal

Decide whether COMMAND_PROTOCOL_V1 is base-cli's cross-language record format or a legacy shim,
and either extend it to carry real operational data or position NDJSON as its successor.

Background

command_protocol.py is the one wire format in the package written deliberately for non-Python
readers. The code says so directly (lib/python/base_cli/command_protocol.py:202-206):

The wire framing is LF-delimited. str.splitlines() also accepts CR, vertical tab, form feed, and
Unicode separators, which would make the Python decoder more permissive than the Bash and Zsh readers.

The framing is well designed for shell consumption: line-delimited, hex-encoded string payloads
(so no quoting or delimiter problems), explicit record counts and end markers, canonical integer
parsing, and NUL rejection.

Two problems limit it as a platform surface.

1. Only two value types exist. FieldSpec.value_type accepts "string" or "boolean"
(_validate_and_store_schema(), line 291). There are no integers, floats, timestamps, lists, or
nested structures. Real operational records need exit codes, durations, byte counts, counts,
timestamps, and tag lists. Every one of those must be encoded as a string by the producer and
re-parsed by each consumer, in each language, with no schema-declared type — which defeats the point
of having a typed schema.

2. Decoding is not streaming, with a high cap. loads_records() takes the whole payload as one
str, does payload.split("\n") to materialize every line, and builds a complete list, with
MAX_RECORD_COUNT = 1_000_000. A large record set is fully resident twice. There is no iterator form.

It is also unpositioned: no page under docs/ explains when to use the command protocol versus
NDJSON, and no reader implementation for another language ships with the repository — so the
Bash/Zsh readers the code is tuned for are not actually verifiable here.

Scope

Choose one and document it:

1. Extend to v2. Add integer, number, timestamp (RFC 3339), and a list type, keeping the
   hex encoding for strings and the LF framing. Add a streaming iter_records(). Publish a JSON
   Schema alongside docs/schemas/v1/command-protocol.schema.json and a reference reader in at
   least one other language (Bash is the natural choice given the existing tuning and the sibling
   base-bash-libs repository).
2. Position NDJSON as the successor. Declare COMMAND_PROTOCOL_V1 frozen and shell-oriented,
   document its exact niche (shell readers that cannot depend on a JSON parser), point everything
   else at the NDJSON contract, and add the migration note to docs/output-contracts.md.

Option 2 is cheaper and may well be right — jq is near-universal now — but the current silence
leaves adopters unable to choose.

Acceptance criteria

• One page states which format to use for which consumer, with the trade-off made explicit.
• If option 1: the new types round-trip through Python and at least one non-Python reader, with
  golden fixtures validated by both, and iter_records() decodes a 1M-record payload without
  materializing it.
• If option 2: command_protocol is documented as frozen with its niche named, and the
  string/boolean limitation is stated rather than implied.
• Either way, docs/schemas/v1/command-protocol.schema.json and the public docs agree with the code.

Validation

Round-trip fixtures in Python plus one other language, consistent with the existing
scripts/validate_contract_fixtures.py / .mjs pattern.

Non-goals

• Do not break the existing v1 framing for current consumers.
• Do not add a binary encoding.