# v0.18: exact reuse and the first replanning fork

Freeze the method, three scientific modules, fixture tests and runners before executing new exposed comparisons. Preserve all 58 prior scientific sources, evidence and the live engine hash. This cycle changes no live controller and draws no fresh seeds.

## Share one waiting projection

The original tail projects the same fitted state under repeated wait actions separately for beds A and B, querying moisture at different travel-plus-lag endpoints. The new implementation advances one cloned state to the longer endpoint and retains each bed's moisture at its requested time. The two scalar answers are the entire memo; it is discarded before the next tail evaluation. There is no global state key, cross-model cache, rounding or persistence. The shared path is bounded to 64 projected ticks; longer requests execute the original two independent projections. Every rollout and all remaining model work still execute.

Keep logical requests identical to the frozen planner's transition counts and use those counts for its original planning and season limits. Separately record executed model transitions and reused requests, with requests = executed + reused. The care wrapper remains the rejected v0.16 two-decision policy. Faster execution alone cannot promote it.

Before exposed work, seed 90101 checks unequal and equal projection endpoints, fallback, non-aliasing, whole forecast and tree equality, legacy charges, physical care equality and budget stopping.

## Fixed benchmark and trajectory parity

Benchmark both retained first-seed v0.14 checkpoints (steady tick 113 and compound tick 115). Run the entire twelve-root search using the original and shared implementation. Two warmup pairs and ten timed pairs per checkpoint; alternate which implementation runs first by pair. Empty temporary memos at every tail invocation. Invoke explicit garbage collection before each search, outside its timer. Require Node's `--expose-gc`. Compare all returned root results exactly after each pair, outside the timer. Warmups and parity executions count as development work.

Report each timing, median paired time ratio, model requests/executions/reuse and sampled heap usage before/after search. Heap deltas include returned results and runtime allocation, can be negative because of GC, and are not peak memory. The memo's structural bound is one working model state plus two retained numbers; report fallback and maximum projection length. Do not claim a total memory reduction from these samples. Ceiling: zero physical actions, ten million executed model transitions and 120 active seconds.

Classify as a useful local speed improvement only if all search and trajectory comparisons are exact, executed transitions fall in both benchmark cases, and median paired shared/original wall time is at most 0.95 in each. Do not retune or rerun to reverse a failed speed criterion. The claim is restricted to these exposed searches on this Mac.

Independently execute the shared implementation for all four retained v0.16 48-tick candidate continuations and the first seed's steady and compound complete candidate seasons. Compare actions, frames, plans where retained, logical work and budget stops against the earlier frozen records. A single new 342-action calibration is shared for complete seasons. Charge 1,494 physical actions total; at most forty million executed model transitions and 120 active seconds. Report logical charges and actual execution separately. No changed-budget care comparison occurs in this cycle.

## Isolate the first forecast/replanning fork

Use the two retained v0.17 checkpoints. For each, choose the original journey and the lowest realized 96-tick cost journey in the earlier v0.17 report, with authored order breaking ties. These four cases are fixed explanations, not a new sample selected from current outcomes.

Reconstruct each 96-tick ordinary continuation. Along exactly the same physical actions, advance a shadow memory that shares the initial checkpoint reading, weather, equipment and applied actions, but receives no subsequent bed-moisture or tank readings. Count its fitting and prediction work. This removes the intervening measurement history, not only the last reading; it is not a deployable sensor policy.

Find the first action differing from the original selected-model script. Retain the most recent new planning decision at or before that action, with its physical checkpoint, current and shadow memories, actual observations, candidate rankings and fitted forecast inputs. Probe the same ordinary menu with current and shadow memories. Also compute the existing heuristic tail from the current selected fitted model. No actual hidden coefficients, future weather or future observations enter these probes.

From that same decision checkpoint, run five continuations to the original window's end:

1. Ordinary replanning, which must reproduce the corresponding baseline suffix exactly.
2. Commit the current ordinary choice once, then resume ensemble care.
3. Commit the shadow's proposed choice once, then resume ensemble care using the actual memory and real readings.
4. Commit the current selected-model heuristic script through its next mechanism operation, then resume ensemble care.
5. Commit the original selected-model script through its next mechanism operation, then resume ensemble care.

Mechanism operations are valve, shade or drain actions. Script commitments are capped at sixteen actions; explicitly record when no operation occurs within that cap. All branches continue to learn from real readings during their commitment. Committing a prefix can prevent ordinary cancellation, so the current-choice arm is a necessary commitment control. Compare shadow/current choices under the same commitment rule; compare the ordinary arm separately. This tests one local journey replacement, not a whole-season policy or a general benefit of ignoring readings.

Report absolute decision and differing-action ticks, local reading histories, model identities/weights, next operations, paired cost/viability/water, and failures or computation limits. Costs and viability are scored over the common remaining suffix within each case; do not average different window lengths into a new generalization claim. Require exact reproduction of the earlier full 96-tick baseline. Ceiling: 2,304 physical actions, sixty million model transitions and 120 active seconds. Reserve the maximum 576 actions per case before running; retain unresolved reservations on failure.

## Delivery

Audit frozen hashes, method ordering, previous evidence, all cost sums and exact outputs. Independently replay one failure-case fork and one control-case fork for verification, charging that work separately. Publish an inspectable read-only view of shared work, timed measurements and paired interventions with exact exports, accessible from the nursery. Update the roadmap using the declared criteria. Software tests, development computation and replay verification remain separate. No result here establishes learned improvement or changes the live default.