feat: integrate explicit CDR recording conversion and fidelity audit

[TomCC7]

Add explicit dimos mem convert for historical MCAP/memory SQLite recordings. Conversion uses an allowlisted retained LCM decoder and the installed generated CDR registry, embeds full dependency schemas, and preserves an adjacent per-row audit. Runtime transports never fall back to legacy decoding.

Parent: #4703. Message Encoding Refactor

This integrates the previously reviewed converter design into the restructured stack with the current registry/value ownership and storage contracts; the prior converter branch and its evidence are preserved.

Changes	Files	Added	Removed
Conversion, batch preflight and existing mem CLI	3	793	0
Fidelity/failure/CLI tests	2	578	0
Self-contained workflow documentation	1	79	0
Installed-package CI	1	7	0

Manual test with an existing trusted local recording (replace old.db with its path, choose new output names):

dimos mem convert old.db review.mcap --dry-run
dimos mem convert old.db review.mcap
dimos mem summary review.mcap

Use .db output to preserve existing vectors; MCAP is recording-only and rejects vectors. For a directory, dimos mem convert ./old-recordings ./cdr-recordings --format mcap --dry-run inventories every candidate without outputs or downloads. Removing --dry-run performs the batch. See the embedded-code walkthrough.

Existing files/reports and competing writers are never overwritten. Originals are read-only. A late single-file payload error publishes no output; partial batch success is explicitly reported with nonzero exit. Pickle, unsupported/custom types, external component stores and attachments are rejected. Header.seq and SQLite IDs/value/tags are audited rather than silently dropped. JPEG bytes are not recompressed. Dry-run validates metadata/schema/counts; full payload integrity is checked during conversion.

Validation:

• 34 source + 34 isolated installed-wheel tests; strict 3 production files; configured scoped hooks passed except unavailable Nix lint.
• Actual installed root CLI dry-run/conversion of already hash-verified cached project go2_mid360_stairs.db: all 4,451 messages, five streams. mem summary reads every stream.
• Actual go2_short.db -> CDR SQLite: all 2,546 messages, four streams, 108 exact 512-dimensional vectors and their observation associations preserved.
• Independent rosbags decoder from embedded schemas compares every mapped field/JPEG byte, source/envelope timestamps, sequences and vector IDs; both source SHA256 values unchanged. Source SHA256: stairs dbb725b763f9d268c47c56e285f5c0150923f488e8fa03b6cdffd26d0dbd125e; go2_short 4f056a042d58ad1dbc8248e408bd61d98bb90e5989cf0520c5400a1088dfe1b9.
• Wheel SHA256 3a152b4b052af8cc66b82f1732ece372d6acfdd15c6906566accdacbc75525fa. Vector tests use official CPU PyTorch only, no model weights or GPU work.

Remaining: dimos mem rerun still assumes legacy to_rerun() methods and skips plain generated values; its external-adapter migration is the next layer. Foxglove local human acceptance/sign-in remains external. Full consumer graphs and legacy dependency retirement remain open. No CI polling, package publication, hardware, history rewrite or merge.