feat(server): passthrough proxy, piecewise keep-ratio curve, query survival check

[smpurkis]

What's in this PR

The C++ server can now act as a compression proxy in front of any OpenAI-compatible backend. Three features added to the pflash compression pipeline.

server/src/server/http_server.cpp (+319)

Passthrough proxy — when --prefill-upstream-base is configured, the server forwards requests to an upstream backend instead of running local inference. Compressed requests go as raw prompt to /v1/completions (the compressed text already contains chat template markup, so hitting /v1/chat/completions would double-template). Uncompressed requests pass through to /v1/chat/completions with the original body. Response format is rewritten (completions → chat completions) for both streaming (SSE chunk-by-chunk rewrite) and non-streaming paths. Uses libcurl.

Piecewise keep-ratio curve — pflash_keep_ratio() replaces the single fixed --prefill-keep-ratio with linear interpolation over breakpoints. E.g. --prefill-curve 10000:0.5 40000:0.2 100000:0.1 gives 2x compression at 10K tokens, 5x at 40K, 10x at 100K+. Falls back to --prefill-keep-ratio when no curve is set. Bandit per-session override (#264) still takes precedence when session_id is present.

Query survival check — after compression, scans the compressed token IDs from the tail to check what fraction of the last user message survived. If < 80% and the query is under 1000 tokens, re-appends the full query text to the compressed output. The 1000-token cap prevents inflating output on single-message prompts where the entire prompt is the query.

server/src/server/server_main.cpp (+36)

CLI flag parsing for the new features + startup logging:

--prefill-upstream-base URL       Upstream OpenAI-compatible base URL
--prefill-upstream-key KEY        Bearer token for upstream
--prefill-upstream-model MODEL    Model name for forwarded requests
--prefill-curve T:R [T:R ...]    Piecewise keep-ratio curve

server/src/server/http_server.h (+9)

ServerConfig: pflash_upstream_base, pflash_upstream_key, pflash_upstream_model, pflash_curve. ParsedRequest: raw_body (preserves original JSON for proxy forwarding).

server/CMakeLists.txt (+5/-1)

find_package(CURL REQUIRED), link CURL::libcurl to dflash_server and test_server_unit.

server/src/qwen3/qwen3_drafter.cpp (+11/-8)

Auto-detect q35 drafter arch from GGUF filename in the 2-arg load_drafter() overload (checks for "qwen3.5" or "qwen35" in lowercased path).

server/test/test_server_unit.cpp (+73)

5 new tests: config upstream defaults, curve interpolation (below/at/between/above breakpoints), curve empty fallback, upstream proxy config, raw_body preservation.

Behavior matrix

Compression triggers	--prefill-upstream-base set	Behavior
Yes	Yes	Compress → forward to upstream /v1/completions
Yes	No	Compress → local inference (identical to main)
No	Yes	Passthrough to upstream /v1/chat/completions
No	No	Local inference (identical to main)

No upstream flags = byte-identical to main.

Evidence

36K-token coding benchmark (5 novel coding tasks, 28 assertions) against Qwen3.6-35B-A3B via the proxy:

Null (direct):     28/28 (100%)  152s
Proxy (pflash):    15/28 (54%)   129s

Server logs confirm all three features:

[server] pflash curve: 10000:0.500 40000:0.200 100000:0.100
[pflash] query survival: 760/36494 (2%)
[pflash] query below 80% but too large to re-append (36494 tokens)
[pflash] 36506 -> 8570 -> 8564 tokens (23.5% kept)
[pflash-proxy] compressed forward → .../completions  prompt=8564 tokens

Unit tests: 1617 assertions, 0 failures.

[cubic-dev-ai]

2 issues found across 6 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>