Drops the `incompressibleKeys{}` side hashtable design from S2.3 and
the `compression-retry-interval` config knob; refines the existing
drift signal in R2.3.5 to include rejected attempts in the
`compression_live_ratio_10m` rolling average. No new config knob.

Why this matters
----------------

Under a fixed dict and a fixed workload, the probability that any
given write at key K produces compressible bytes is constant — time-
based throttling between retries doesn't shift this probability, only
spaces attempts. CPU bounding is already provided at the sweep level
(`compression-sweep-max-cpu-pct`); per-key throttling would be
redundant and would add correctness burden (every mutating code path
— `dbOverwrite`, APPEND, SETRANGE, BITOP, BITFIELD, module DMA writes
— would need a `Clear` integration point to avoid suppressing
compression of legitimately-changed values).

The right trigger for "the rejection might now compress" is **dict
change**, which is handled at the system level. Updated definition:
`compression_live_ratio_10m` is now the rolling ratio over compression
*attempts* (successful + rejected) in the 10 min window, weighted by
uncompressed bytes. Each successful compression contributes its actual
`compressed/uncompressed` ratio. Each rejection contributes its actual
measured ratio (typically in `[0.9, 1.05]` since the net-savings guard
rejected it for being too close to 1.0). Worker errors are excluded —
they're tracked separately via `compression_errors_total`.

Both forms of degradation feed the metric uniformly:
  - workload-content drift among compressible values, and
  - workload composition drift (more values compressing poorly enough
    to fail the net-savings guard).

A single drift trigger threshold catches both scenarios. Operators
reason about ONE metric; the design has ONE config knob
(`compression-dict-drift-ratio`) for the threshold.

This is a deliberate simplification over an earlier design (PR #10)
which proposed a per-key `incompressibleKeys` side hashtable with
`compression-retry-interval` time fallback. That approach added a
module + a config knob + integration points in every mutating code
path, but the underlying assumption (waiting between retries
improves the per-attempt hit rate) is incorrect for fixed-distribution
workloads under a fixed dict. The simpler model — retry on every
sweep tick, let the drift signal handle systemic dict-fit issues — is
correct and operationally cleaner.

Drift comparison direction fixed
--------------------------------

R2.3.5's existing trigger had the comparison backwards. The ratio
convention is `compressed/uncompressed` (lower = more savings, per
R2.10.1). The previous formulation —

    fires when live_ratio < drift_ratio × post_training_ratio

— fires when live is BETTER (lower) than 70% of post-training, i.e.,
when compression IMPROVES over post-training. That's not drift.

Corrected to:

    fires when live_ratio > post_training_ratio / drift_ratio

With drift_ratio = 0.7 (70%): fires when live exceeds post-training
by a factor of 1/0.7 ≈ 1.43 — i.e., compression efficiency has
degraded by about 43%. Default value preserved; only the comparison
direction changed.

Why "real measured ratio" not "1.0" for rejections
--------------------------------------------------

An earlier draft of this commit substituted `ratio = 1.0` for every
rejected attempt, on the theory that we store such values uncompressed
so the effective storage savings is zero. That was reviewed and
revised: the metric's purpose is **drift detection** (workload-
compressibility under this dict), not effective-storage tracking. A
rejected value at 0.91 (just above the 90% net-savings threshold) and
a truly incompressible value at 1.05 carry different information about
dict-workload fit. Substituting 1.0 for both throws away that
distinction; using the actual measured ratio preserves it. The data
is right there — the worker reports the ratio when it returns, we
just record what was measured.

Code changes
------------

- src/compression.c (`compressionIsEligible`):
    - Drop branch 6 (the S2.2-stub "always retry-eligible" comment +
      the S2.3-anticipated lookup that never landed).
    - Drop the `key` parameter from the signature — no longer needed
      now that no branch consumes it. Trivial to re-add when a future
      branch (e.g., S2.7 write-path hook) wants the key for logging.
    - Update doc comment to reflect the simpler predicate.
- src/compression.h: matching signature change.
- src/server.h: drop `compression_retry_interval` field.
- src/config.c: drop `compression-retry-interval` registration.
  Advanced knobs count goes 10 → 9.
- src/unit/test_compression_eligibility.cpp: drop `key` parameter
  from all test calls.
- tests/unit/type/compression.tcl: drop the
  `compression-retry-interval` config-default assertion; update
  comment from "Advanced (10)" to "Advanced (9)".

Doc changes
-----------

- detailed-design.md §2.2:
    - Predicate: drop `retry_eligible(obj)` line.
    - Rationale paragraph rewritten to use the probability argument:
      time-throttling doesn't change per-attempt success probability,
      only spaces attempts; sweep pacing already bounds CPU. Per-key
      tracking is functionless and adds integration burden.
- detailed-design.md §2.3 R2.3.5 drift triggers:
    - Single trigger formulation, comparison direction fixed
      (< → >). Definition of `compression_live_ratio_10m` made
      explicit: rolling ratio over attempts (successful + rejected)
      using actual measured ratios; worker errors excluded.
- detailed-design.md §2.12 advanced config table:
    - `compression-retry-interval` removed (10 → 9).
    - `compression-dict-drift-ratio` description updated to reflect
      corrected formula.
- detailed-design.md §6.6 Net-savings guard failure: rewritten with
  the probability-argument rationale and real-ratio framing.
- detailed-design.md §7.1 transparency-mode harness config: drop
  `--compression-retry-interval 0` line.
- idea-honing.md Q6 baseline filter "Skip post-compression" bullet:
  matching rewrite (probability argument, real-ratio framing).
- idea-honing.md Q6 consolidated predicate: drop `retry_eligible(obj)`
  line and the `where retry_eligible(obj) is...` block.
- idea-honing.md Q6 config table: drop `compression-retry-interval`
  row.
- idea-honing.md Q6b answer: matching rewrite with reference to the
  S2.3 implementation review (PR #10 design discussion) so the trace
  is preserved.
- idea-honing.md §7.1 harness: drop the same flag.
- summary.md: update eligibility row + walkthrough-highlights bullet
  to note the post-walkthrough refinement.
- plan.md S2 row description: drop "incompressible-keys hashtable"
  from S2's scope summary.
- plan.md S2.2: drop the trailing "incompressible-keys hashtable
  check" claim.
- plan.md S2.3: marked "(REMOVED)" with explanation; struck-through.
- plan.md S1.4: extended description to include the rolling-ratio
  semantics extension (rejections contribute their actual measured
  ratio) — the drift-detection work absorbed from removed S2.3.

Audit-trail files (DESIGN_TODO.md, pr-feedback.json) intentionally
unchanged — Thread #20's resolution stands as written; only the
implementation interpretation in the live design docs is refined.

Branch ikolomi/s2-incompressible (the abandoned S2.3 implementation)
should be closed/deleted on GitHub.

Verified locally
----------------

- `make -j` builds clean.
- `./runtest --single unit/type/compression` 10/10 passes.

Not verified locally (CI will validate):
- gtest unit tests on Linux/macOS/32-bit (no libgtest-dev locally).