…ase/bold/dedup)

…ERY_MEDIA_EXTS

…sage loop

…play stripping

…ousResearch#34178)

The non-streaming dispatch path used a blind re.sub(r"MEDIA:\s*\S+") to
scrub residual MEDIA: tags from outbound text. extract_media already strips
recognized tags upstream, so the blind pass only erased unknown-extension
tags — silently hiding the model's intent to deliver a (non-deliverable)
file (NousResearch#32644) and diverging from the streaming display path, which was moved
to the shared _MEDIA_TAG_RE. Route this 5th site through the same object:
defensive no-op for recognized tags, unknown-ext tags now left visible and
consistent across streaming/non-streaming. Closes the last blind MEDIA regex.

…xt leak (NousResearch#28989, NousResearch#24032)

extract_media/_TOOL_MEDIA_RE now recognize Windows drive-letter (C:\...,
C:/...) and UNC (\\server\share\...) paths — incl. spaced and bold forms —
so the MEDIA: tag is stripped from the user-visible message instead of leaking
verbatim as plain text (the headline symptom of NousResearch#28989 and NousResearch#24032).

Safe native-Windows *delivery* is deliberately NOT enabled here: a Windows
credential denylist is unwinnable (\\?\ prefixes, 8.3 names, ADS, UNC all
defeat prefix matching), so it requires the allowlist-model path-validation
security PR (the L0 follow-up, GRILME Q7a). validate_media_delivery_path
fail-closes drive-letter/UNC paths (\w:[\\/]|\\\\ — \w catches Unicode
homoglyph drives) on every host, so they are dropped + surfaced via
media_dropped rather than delivered. This also closes a latent fail-open for
quoted Windows paths on native Windows hosts, where the POSIX credential
denylist matches nothing.

_LOCAL_PATH_RE (bare-path auto-detect) is intentionally left POSIX-only to
avoid false positives on 'C:' in prose.

…osed guard

- Move media_dropped success + send-error-guard assertions out of
  test_send_message_tool.py (gated by importorskip('telegram'), which neither
  [dev] nor [all] installs, so they were silently skipped in CI) into a new
  telegram-free test_send_message_media_dropped.py. Adds an end-to-end
  Windows-tag strip+drop test through the send_message tool path.
- Strengthen the Windows-not-deliverable test: force is_absolute()=True (the
  native-Windows condition) so the fail-closed guard ALONE is what rejects
  drive-letter/UNC/homoglyph paths, instead of the incidental POSIX reject
  masking a deleted guard.

Address code-review test-quality findings (behavior over implementation):
- Extract tool_media_paths() so run.py's two MEDIA-salvage sites share one
  tested helper instead of duplicating group(1).strip().rstrip('",}'); replace
  the impl-coupled _TOOL_MEDIA_RE direct tests (which had drifted from the
  caller's salvage) with a behavioral test of the helper.
- Replace the stream-consumer _MEDIA_RE-is-_MEDIA_TAG_RE identity assert (a
  change-detector) with a behavioral equivalence test: streaming display-strip
  output must equal extract_media's cleaning across inputs — the real zero-drift
  contract, surviving a refactor that doesn't literally share one object.

…ation, trim what-comments

Concision pass: inline the one-use _media_ext_alternation() into the
_MEDIA_EXT_ALT constant; drop the redundant non-capturing group in the
Windows guard; condense the frozenset header + guard comments to the
load-bearing why (single-source + exclusion rationale; fail-closed +
homoglyph rationale), dropping the what-enumeration and an opaque
internal cross-reference. No behavior change.

test_windows_media_path_partitions_to_dropped asserted the composition of two
already-covered behaviors (partition split + validate Windows->None) and is
also covered end-to-end by test_windows_media_tag_stripped_and_dropped_end_to_end.

…havior-preservation

…ensions

# Conflicts:
#	gateway/platforms/base.py