compress: route every trigger path through compress_and_save_session ...

- CompressionWorker keeps the real-token gate (ctx.context_tokens) but now
  delegates to compress_and_save_session(reason="postturn") — it used to
  call compress_context directly with its own save logic and silently
  no-op'ed on any summarizer failure, leaving the session over the
  threshold until the next turn's gates
- pre-turn and /compact gates use real_baseline_estimate instead of the
  chars//3 heuristic (mid-turn already did), so code-heavy contexts
  compress in time instead of tripping ContextTooLargeError
- summarizer input keeps head + tail instead of head-only truncation,
  which silently dropped the newest summarized messages closest to the
  keep window
- critical tool results over the 4000-char budget keep head+tail halves
  instead of collapsing to the 800-char non-critical preview (a 3999-char
  read survived whole, a 4001-char one lost 98%)
- images count 500 tokens each, not 500 per message carrying them
- _plan_compression is the single partition decision shared by
  compress_context and would_compress, so the dry-run prediction cannot
  drift from the real path; the trigger reason is now logged
  (preturn/midturn/postturn/forced)

config: raise context compression threshold 0.70 -> 0.90 ...

Midturn compression now splits the in-flight turn correctly, so it's safe to
let the context grow closer to the limit before compressing. 0.90 gives the
model more raw context per turn and trips the compressor less often.

Worker test context_tokens bumped 50_000 -> 60_000 (76% -> 91.5%) to stay
above the new threshold; other compression tests pass the threshold as an
explicit argument and are unaffected.

Co-Authored-By: Claude <noreply@anthropic.com>

compression: fix auto-compress no-op on few-huge-messages + honest status ...

Root cause: the compression gate (should_compress) measures tokens, but the
partition measures message/turn count, and CompressionStarted was emitted
before the attempt. For navi_code's "few very large messages" shape (one big
file read = 1 user + assistant + 1 huge tool result, 66k tokens in 3 messages)
the gate fired, the UI showed "compression", but partition returned
to_summarize=[] -> compress_context None -> nothing shrank. The agent kept
going until the window overflowed. It wasn't running "during" compression —
there was no compression, just a no-op the user mistook for one.

A. Per-message head/tail truncation in context_builder.build(): oversized
   tool/assistant messages (over context_message_token_budget, 0=num_ctx//6)
   are capped head+marker+tail in the LLM view only (model_copy — stored
   history and reloads are never affected). A single huge tool result can no
   longer alone blow the window; user/system messages are never truncated.

B. Token-budget hard-truncate fallback in compress_session: when partition
   no-ops but tokens exceed the threshold, drop oldest turns to num_ctx*0.5.
   _hard_truncate is now token-aware (was a fixed message-count floor that
   no-oped on <=6 messages even when huge). New would_compress() predicts
   compress_session's real outcome with no LLM call.

C. Honest CompressionStarted: _compression_events_midturn/_preturn emit it
   only after would_compress() confirms the partition (or token-budget
   fallback) can actually shrink the stored context — no more "compression"
   status with no ContextCompressed to follow.

Bonus: post-turn CompressionWorker now passes keep_recent_messages=
max(12, context_keep_recent*2), matching the midturn path, so a single long
autonomous turn compresses post-turn too (was always a no-op).

Tests (+14): would_compress agreement, token-budget fallback, token-aware
hard_truncate, build() truncation (preserves user, no mutation, head+tail),
agent no-CompressionStarted-when-nothing-to-compress, worker single-long-turn.

Co-Authored-By: Claude <noreply@anthropic.com>