[terrelln]

Summary

This PR fixes dictionary lifetime and selection behavior in the experimental ZSTD_d_refMultipleDDicts API. A populated DDict set now acts as an explicit mapping from frame dictionary IDs to borrowed DDicts, with no implicit default or remembered selection. The separate hash-table lookup and growth fixes are part of PR #4927.

The four commits are:

1. 8a1bf5f — Clear remaining multi-DDict references.
   Centralizes set cleanup and releases the set when loading an empty dictionary or disabling multi-DDict mode. This makes the documented “return to no-dictionary mode” behavior reliable and lets callers immediately free every formerly registered DDict.

2. 201aae6 — Make local dictionaries and DDict sets mutually exclusive.
   Loading a dictionary or referencing a prefix discards the set; registering a DDict discards any local dictionary or prefix. This gives dictionary replacement an unambiguous lifetime and prevents two competing sources of dictionary state from coexisting.

3. 845b46b — Require exact matches for nonzero frame IDs.
   Selects dictionaries only after a complete, valid ordinary-frame header is available, before installing decoding state. Missing nonzero IDs return ZSTD_error_dictionary_wrong instead of falling back to the previous dictionary. ID-less frames use a registered ID-0 DDict when present, or no dictionary otherwise.

4. 82e0df9 — Resolve set entries independently for each frame.
   Keeps the single-dictionary state empty while a set exists and resolves each frame into a local DDict pointer. Explicit-dictionary APIs remain independent of the set, tracing no longer needs a retained DDict pointer, and implicit set-based legacy decoding is rejected.

   The default DDict was an extremely confusing concept, as it ended up being either the last DDict loaded, or the last DDict used regardless of load order. Also explicitly disables legacy frame decoding when there is a non-empty DDict set. The prior behavior was 100% not correct, so this isn't a big loss.

Design rationale

• Nonzero IDs require exact matches. Falling back to the last registered or selected dictionary makes decoding depend on unrelated earlier operations. ID-less frames instead use a registered ID-0 DDict when present, or no dictionary otherwise. This allows ordinary dictionary-free frames without requiring an empty DDict registration, while preserving explicit support for raw-content dictionaries.
• The parameter remains sticky. Loading a dictionary or prefix replaces the active set without disabling ZSTD_d_refMultipleDDicts. The parameter controls subsequent refDDict() registrations; changing dictionary storage does not silently change configuration. Existing single-dictionary and single-use-prefix behavior is preserved.
• Explicit arguments take precedence. decompress_usingDict(), decompress_usingDDict(), and the bufferless APIs must use the dictionary supplied by their caller, even if the context also contains a set. Legacy decoding remains available through these explicit or single-dictionary paths rather than inheriting an arbitrary set member.
• Header-parser optimization is left separate. The one-shot set path currently parses the header before dictionary initialization, then uses the established frame decoder. This keeps initialization ordering and existing error handling clear; eliminating the redundant parse can be a focused follow-up.

Testing

Adds regression coverage for immediate dictionary release, replacement in both directions, missing nonzero IDs, registered and absent ID-0 entries, dictionary-free frames mixed with dictionary-backed frames, concatenated and magicless input, fragmented streaming and its one-shot shortcut, explicit-dictionary APIs, prefix compatibility, and the legacy-format boundary. An ID-less frame that requires dictionary bytes also verifies that the fallback does not reuse a previous dictionary.

Passed:

make -C tests -j8 fuzzer legacy
tests/fuzzer -i0 -s0
tests/legacy