Roll back failed override batches and keep structural-reform nodes across parameter rebuilds #2374

GPT-6.1 Sol's round-3 review of #2305 approved it, with four low-severity notes, all reproduced against #2305's head (603fecb):

1. A failed batch of legacy overrides stays partly applied. Scenario.from_reform({rate: {"2027": 0.3}, "gov.not_a_real_parameter": {"2027": 1}}).apply(sim) raises, but the rate already reads 0.3 for 2027, and nothing records it. The next rebuild silently reverts it.
2. Nodes a structural reform adds don't survive a rebuild or reprocessing. After Simulation(reform=(add_bonus, {"test_budget.bonus": 7})), both sim.apply_parameter_changes(...) and no_economic_assumptions.apply(sim) fail with Could not find the parameter test_budget.bonus. process_parameters has no rollback, so the failure also leaves a half-processed tree.
3. Two shared caches are handed out uncopied: a memo entry's projected timelines, and the module function projected_timelines(). Nothing in the repo mutates them, but the docs said they were copies.
4. The docs don't say that Scenario.apply on a built simulation edits the annualised values, nor that single-day keys and a scenario's own modifier run after the UC rebalancing and parameter-gated structural reforms.

Found while reviewing #2305 (Budget CI W1).