v0.7.65: email template alignment, logs enrichment, execution files unique keys - #6481

Execution file keys were built as
execution/{workspaceId}/{workflowId}/{executionId}/{fileName}, so two files
sharing a display name within one execution resolved to the same key and URL.
The second upload overwrote the first in object storage and updated its
workspace_files row instead of inserting, silently losing a file. Any trigger
that ingests multiple attachments hits this — repeated screenshot names, mail
clients that reuse inline-image names, or a loop emitting the same output name.

generateUniqueExecutionFileKey now allocates a unique directory segment
(.../{executionId}/{uuid}/{fileName}) and uploadExecutionFile uses it, so every
execution file gets its own key. The uniquifier is its own path segment rather
than a filename prefix because presigned URLs carry no content-disposition: the
key's final segment is the name a consumer sees, and a prefix would rename
every download.

The deterministic generator is renamed to generateLargeValuePayloadKey and
takes the payload id instead of a free-form file name, so no user-supplied name
can reach a key without a uniquifier. Its output is unchanged — determinism is
load-bearing there, since the cleanup job matches those keys by LIKE pattern
and the trace store recovers workflowId by segment position.

Every reader tolerates the extra segment: key parsers use parts.length >= 5
with fixed indices, storage providers write a preserved key verbatim, and
local-disk storage already creates the dirname recursively.

…design system (#6479)

* improvement(emails): align the email design tokens with the platform design system

* improvement(emails): preview every template in a single gallery page

* improvement(emails): align font stack, logo and social-icon sizing with the platform

* improvement(emails): tokenize the CTA and footnote, and enforce the platform mirror with a test

* fix(emails): restore row spacing in the payment-failed details box

* fix(emails): restore row spacing in the batch-invitation workspace list

…ete (#6478)

Incompleteness is one-way: once any guard trips, every later model projection in
the run fails and the user is left with a single opaque sentence. Every guard
could set it and none recorded which, in a file that imported no logger at all,
so the cause could not be recovered after the fact.

Name each guard with a static reason literal. Originating causes log at error
because they permanently fail the run and error is the only level that survives
every default the logger falls back to; reasons that merely carry an upstream
fault forward log at warn so one fault does not read as several. The decrypt
catch no longer discards its cause.

No behaviour change. Reasons are static literals and the logged input path is
block/field names; no resolved value is recorded.