Follow-up to the delivery-layer fix in c2b8de4. Appends a shared
_MEDIA_GROUP_RULES block to the telegram, whatsapp, discord, slack,
signal, and bluebubbles platform hints so non-Anthropic models adopt
the new response shape reliably:

- Example-first (small models imitate the shape directly)
- Explicit "2 or more ... ALWAYS" trigger
- Rules call out the split/merge pitfalls (blank line ends album,
  no prose between MEDIA: lines, trailing single caption shares)
- Points at the telegram-media-group-captions skill for complex cases

weixin/wecom intentionally skipped — native album semantics there
differ from Telegram/Signal and warrant a separate pass.

Manually verified on a live Telegram deployment with two models:
  - gpt-5.4 adopted the album shape without any skill hint
  - gemma-4-26b (heretic-apex) adopted it once the rules landed;
    previously needed an explicit skill-name hint in the user turn

Tests: 92 new cases in TestMediaGroupRules, full test_prompt_builder
suite 150/150 green.

…treaming path

The feature added _parse_content_blocks + send_media_group to BasePlatformAdapter,
and wired the streaming already_sent path through them, but left
_process_message_background using the legacy extract_media + extract_images +
per-item send_image_file loop. On platforms where streaming is off (or the
response arrived without streaming), a response containing MEDIA:/caption pairs
produced one text blob with all captions and N individual sendPhoto calls —
never a real Telegram album.

Detect MediaGroupBlock in the parsed response and dispatch TextBlock /
ImageBlock / MediaGroupBlock directly, short-circuiting the legacy pipeline
only for album-shaped responses so TTS/auto-TTS/local-file fallbacks for
plain-text replies are untouched.

…s_message_background

Covers the non-streaming delivery path that bypassed _parse_content_blocks
and produced a caption-blob + N individual sendPhoto calls instead of a
single sendMediaGroup. Three cases:

- 3-item album dispatches one send_media_group with per-item captions;
  no send_image_file or send_image calls; no caption text dumped to send()
- 12-item response splits into 10 + 2 albums (Telegram's group cap)
- Prose before MEDIA: lines is sent as a text message with MEDIA: tokens
  stripped; album delivered separately

Guards against future upstream refactors that reintroduce a bypass of the
content-block dispatcher.

…d-media-group skill

_MEDIA_GROUP_RULES pointed at 'telegram-media-group-captions', but no
such skill file exists in the tree. Claude/GPT models still worked
because they learn directly from the inline few-shot example, but
smaller / open-weight models (gemma4, etc.) tried to look up the named
skill in Hermes' auto-injected skill listing, found nothing, and
degraded to needing the skill name spelled in the user turn.

Since album delivery now spans five platforms (Telegram, Discord,
Signal, Slack, Feishu), rename the referenced skill to the
platform-neutral 'send-media-group' and ship the actual SKILL.md file
with it.

The new skill's description carries keyword triggers (album, media
group, gallery, photo collection, multi-image, per-item captions,
every platform name) so the skill index surfaces it whenever a model
is about to deliver 2+ images. Body covers the exact MEDIA: response
shape, all edge cases (auto-split at 10, shared vs per-item captions,
blank-line separators), and how each of the 5 native-grouping
platforms renders the result.

Minimum-footprint fix:
  - 1 new SKILL.md file
  - 1 string replaced in agent/prompt_builder.py
  - 1 test assertion updated

No change to adapter code, dispatcher, parser, or other skills.