fix: restore implicit-cache hit rate — per-conversation session IDs + persistent signature cache - #367

Signatures are replayed verbatim on every subsequent turn of a
conversation. The 2h TTL silently degraded them: tool_use parts fell
back to the skip marker and thinking blocks were dropped entirely once
their family entry expired, rewriting conversation history mid-session
and busting the upstream implicit-cache prefix from that point on. A
proxy restart had the same effect because the caches were in-memory
only.

Signatures are now kept for the whole process lifetime and persisted to
~/.config/antigravity-proxy/signature-store.json (atomic tmp+rename,
mode 0600, debounced flush plus periodic/exit safety flushes, override
with ANTIGRAVITY_SIGNATURE_STORE). The store loads at boot, so restarts
no longer rewrite long-running conversations. All persistence failures
fail open — store I/O can never break request forwarding.

/test/clear-signature-cache now clears both caches and removes the
persisted store. Cross-model safety is unchanged: the family lookup
still gates Gemini replays, only the artificial expiry is gone.

Google streams usageMetadata progressively: early chunks carry an
interim promptTokenCount and no cachedContentTokenCount yet, so
message_start latches input_tokens = (nearly-full prompt) - 0 whenever
the request is served mostly from cache. message_delta then corrected
cache_read_input_tokens but never input_tokens, leaving clients that
combine the two (e.g. gateways computing total = input + cache_read)
with a phantom ~2x prompt and a fabricated ~50% cache-hit rate.

Replay-proof against a captured 136k-token conversation: same body,
stream:true reported input=134,572 + cache=130,467 (phantom 265k total,
'49% hit') while stream:false reported the true 136,389 total with
130,467 cached (95.8% hit). The final usageMetadata chunk is
authoritative, so message_delta now also carries
input_tokens = promptTokenCount - cachedContentTokenCount (same formula
the non-streaming path already uses; Math.max guards odd backends).
Verified end-to-end after the fix: warm repeats report ~5.7k input +
~126k cached (95.7% hit) and the next turn ~5.4k + ~130k (96.0%).

…tch (Gemini only)

The synthetic-message injection ('[N tool executions completed.]' +
'[Continue]', and the mid-conversation splice for interrupted tools)
was added wholesale in 426acc4 with no issue reference and no recorded
upstream error. A/B replay against the live Cloud Code backend shows
Google accepts raw mid-tool-loop, parallel-tool, and interrupted-tool
histories natively and answers identically without the injection, so
for Gemini targets the fabrication is pure downside: a fake user
'[Continue]' becomes the model's most recent instruction on nearly
every adaptive-thinking agentic request, and the toolResultCount-bearing
text differs between requests.

Claude targets keep the recovery because the Anthropic API
hard-rejects conversations with unclosed tool_use ids, so the
interrupted-tool synthesis may still be required there.

Anthropic semantics bill all generated tokens (thinking included) as
output_tokens; the backend reports internal reasoning separately as
thoughtsTokenCount and bills it as output quota. Mapping only
candidatesTokenCount silently undercounts thinking-heavy responses by
40-60% (measured: determinant task reported 1411 while generating ~2400
tokens). Fix both the streaming and non-streaming usage paths.