feat: weight edits read the same on every surface, and experiments record their checkpoint - #32

Transport.enqueue(), the client's enqueue() and emit() return true when
the item was queued and false when it was dropped or the SDK is off or
shut down. Nothing reads the result yet; weightsEdit() will.

…ord the same weight edit as Python

weightsEdit() sends one notification event with payload.kind
"weights.edit", with the same keys, values, validation and session
shape as the Python SDK's weights_edit(). It never throws and returns
true only when the event was queued. weightsDelta() computes shape,
changed, l2 and maxAbs for nested arrays and typed arrays.

packages/sdk/test/fixtures/weights-edit.json holds the events an edit
must send, recorded from the Python SDK. Both test suites read it.

… or leaves through the otlp sink

A span or a log record named weights.edit with coding_sh.weights.*
attributes is now read as the weights.edit notification the SDK sends,
so the session gets the same stat, tag and line. Before, a span became
a generic span and a root span lost its attributes.

The otlp sink and serve's forward send one weights.edit span per edit
with the same attributes, and the invoke_agent span carries
coding_sh.stats.weight_edits and coding_sh.stats.weights_changed.
Metadata mode keeps the counts and the norms: the forward allowlist
drops the tensor name, shape, hashes and note.

…nd an impossible shape or count is not printed as a total

MCP's session_get dropped an edit whose payload had no well-formed
field. The line no longer says "3 of 0 weights" when the count does not
fit in the shape, reads no shape of more than 16 dimensions (the SDK's
limit), and never rounds a norm to Infinity. The docs now say which
edits serve's forward sends.

…g the tool surface

The words added for weight edits are paid for by a shorter opening and
a shorter note on the turn argument, so the description budget still
holds when other tools add words.

…nd exp compare names both

exp run --weights <SHA256> stores metadata.weights = { sha256 } (64 hex, lowercased; coding.sh
hashes nothing). exp compare shows each side's stamp in the normal and the --given header, warns
when both sides ran the same weights and neither regrades the other, and says so when only one
side is stamped. exp show, the exp run summary, the viewer and the MCP tools experiments_list and
experiment_compare carry the same facts.

--arm takes one stamp per arm (--weights base=<SHA> --weights edited=<SHA>). --watch and matrix
specs refuse the flag. --resume keeps the stored stamp and refuses a different one; the stamp is
part of the job config. A regrade inherits the stamp.

…, not 0.000

A count or score below 0.0005 (a gain per edited weight of 0.0001) printed as 0.000 and read as
zero. It now prints two significant digits; values that three decimals already show are unchanged.

The Metric A/B guide names three keys for an arm that reports a model edit (edited_weights,
gain_pp, gain_per_weight), says the two gain keys need --higher, and shows the table. A CLI test
runs a stub arm that prints them and checks the three rows and their direction.

…ts and weights.changed

The two --where keys had SQL only in the SQLite store, so a query that
used them was refused on the other two.

…LP, in adapters

l2=delta.norm() in torch is a 0-d tensor, not a Python number, so the
edit was refused and dropped with a warning while the docs said torch
scalars count. A value with ndim 0 and item() is now read through
item(); a tensor of more than one value is still refused.

# Conflicts:
#	CHANGELOG.md
#	docs/reference/api.md

# Conflicts:
#	CHANGELOG.md

# Conflicts:
#	CHANGELOG.md

# Conflicts:
#	CHANGELOG.md