[wannagoxuf]

Problem

The snapshot API contract (include/strata/core/conversation_snapshot.hpp) says:

A nonzero unchanged_tokens is valid only for storage retained from an image actually restored into this
session, bounded by every subsequent rewrite. Equal token IDs alone do not establish that its K/V bytes
are unchanged.

The bookkeeping half of that - which bytes a capture claims are unchanged, and where the claim stops - is
covered in conversation_cache_test.cpp (retain / limit_reuse / take_reuse). The half that is not
covered is the byte half: whether the bytes really are unchanged.

That question is live in the two pull requests that touch the reuse path (#3711 and the harvest discussed in
#3321), and it is decidable in the shapes the cache actually produces.

Approach

A new test (src/core/conversation_reuse_identity_test.cpp, 21 checks) that covers:

• a tail rewrite leaves the prefix bytes alone, and changes the tail bytes. This is what the
  unchanged_tokens claim rests on: the bytes before the first rewrite are still the ones the session holds.
• a different prefix gives different bytes - the negative control. Without it the first case would prove
  nothing, since identical bytes could just mean the comparison is too coarse to see a difference.
• the claim's bound: limit_reuse cuts it at the first rewrite, a rewrite from the start drops the whole
  donation, and a later continuation cannot widen a claim that was already cut.
• the claim survives the capture: with a nonzero unchanged_tokens the capture keeps the donated prefix
  bytes and not the device's, and past the claimed prefix it takes the device's bytes.

The engine-level half - two independent computations of the same token prefix through different chunkings -
needs a real model and is not decided here. It is measured with STRATA_STATE_HASH=1 (the engine's
per-request fingerprint of the session's K/V) and the result is in #3321: on 4x4090 with
Qwen3.8-Flash-Next IQ3_XXS, the same prompt through --prefill 4096 and --prefill 1024 gives different
main-K/V fingerprints (2b0d5053ac251873 vs 42e1a97e3930a0fe), while the same chunking twice gives the
same fingerprint. So equal token IDs do not, in fact, establish equal bytes across a chunking boundary - the
contract's caution is observable, not theoretical.

Validation

Passed:

• Build on origin/main (7dd5cf1): cmake -DCMAKE_BUILD_TYPE=Release -DSTRATA_ENABLE_CUDA=ON -DSTRATA_BUILD_CONVERSATION_TESTS=ON + cmake --build . -j 4 (CUDA 12.8, driver 580.178.04, 4x RTX 4090).
• ctest -R conversation --output-on-failure -j 2: 8/8 passed, including the new
  conversation_reuse_identity_test (21 checks).
• No production code is touched: the diff is one new test file and its add_test registration.

Not run:

• The HIP and SYCL backends are not built here (no toolchain), so the test is not exercised on them. The
  test uses only conversation_kv_save / conversation_kv_verify / ConversationCache, none of which are
  backend-specific, but that is an argument, not a run.
• The engine-level identity measurement (the STRATA_STATE_HASH=1 runs) is on the CUDA engine only.

Failed: nothing.

Limits

• This test decides the bookkeeping and the byte-level properties of the shapes the cache produces. It does
  not decide whether two independent computations of the same prefix agree - that needs a model, and the
  answer there is no (see the measurement above).
• The fixture writes the K/V bytes itself, so it can show that a rewrite leaves a prefix alone and that a
  different prefix differs; it cannot show that a real model's prefix bytes are stable across chunkings.

Related

• [Engine] reuse warm pages for slot copies and superseded parks #3711 (the page-reuse part of the same series)
• [Engine] one thread per GPU for the per-stage copies (park capture and slot shuttle) #3742 (the per-stage concurrent copies; no overlap with this test)
• [Engine] Slot <-> pool copy bandwidth is ~5 GB/s (takes/gave back/park), ~19 GiB conversation costs ~11 s per agent turn; PCIe probe says 26.8 GB/s #3321 (the discussion this test comes out of)