@@ -1,33 +1,33 @@
-phase,method,workload,mean_ns,standard_error_ns,confidence99_halfwidth_ns,samples,allocated_bytes
-Compilation,SystemCompile,Arithmetic,21057.400173611113,480.06713333032013,2596.146657766782,8,4447
-Compilation,FastCompile,Arithmetic,2136.472993116752,29.83925431808885,161.36718177467185,8,943
-Compilation,LightCompile,Arithmetic,1861.2931854812869,14.306173933474842,85.2478605594403,7,943
-Compilation,SystemCompile,Array,46886.166581483645,561.0764982423005,3034.2357860623824,8,5120
-Compilation,FastCompile,Array,2750.989670356446,24.45831453192665,145.74260011322139,7,1151
-Compilation,LightCompile,Array,2768.623869131542,29.66190207535884,160.40808168169767,8,1151
-Compilation,SystemCompile,Switch,29992.961711711712,233.24966158584692,1261.3867672017197,8,6182
-Compilation,FastCompile,Switch,2425.097851613464,42.57471854073646,230.23907610197057,8,1343
-Compilation,LightCompile,Switch,2286.3197476641903,10.490214069947784,56.729845278147806,8,1360
-Compilation,SystemCompile,CapturedConstant,34270.82298108068,315.7630793153843,1707.60963643542,8,4904
-Compilation,FastCompile,CapturedConstant,2065.5865263146625,23.705268182687245,128.19530538784718,8,1183
-Compilation,LightCompile,CapturedConstant,1975.462104590091,44.4838550235361,240.56346191163306,8,1183
-Construction,SystemTree,Arithmetic,440.8064759139369,1.9679978125448743,10.6427009657735,8,584
-Construction,LightTree,Arithmetic,178.25360727976607,0.9876112338033758,5.340885525790328,8,360
-Construction,SystemTree,Array,1058.828216831608,18.78200639040111,111.91852337604925,7,1192
-Construction,LightTree,Array,401.1927925192827,3.773679071513119,20.40761307909032,8,672
-Construction,SystemTree,Switch,1126.56504952069,10.951442884606463,65.25763494089202,7,1432
-Construction,LightTree,Switch,209.80873104222894,1.3395839536704643,7.2443126443456105,8,824
-Construction,SystemTree,CapturedConstant,508.16101190874906,4.123871927411091,22.301414902402737,8,528
-Construction,LightTree,CapturedConstant,237.72549601577592,0.8994772171331511,4.864267117805024,8,272
-Invocation,SystemInvoke,Arithmetic,1.1857331990510527,0.00393391632142369,0.021274157301601867,8,0
-Invocation,FastInvoke,Arithmetic,1.0989445437726775,0.00231112943902114,0.012498316489941222,8,0
-Invocation,LightInvoke,Arithmetic,1.0980764832439314,0.002266402297427581,0.012256437362840716,8,0
-Invocation,SystemInvoke,Array,1.0417457639223606,0.0023781178586783683,0.012860581993508882,8,0
-Invocation,FastInvoke,Array,1.2857391691880478,0.0006201253789979213,0.0036952131355322184,7,0
-Invocation,LightInvoke,Array,1.2984493339059076,0.004658099760368262,0.025190456260840082,8,0
-Invocation,SystemInvoke,Switch,1.0546706028009836,0.0016552381833178306,0.008951333634564273,8,0
-Invocation,FastInvoke,Switch,1.3091380151624001,0.0036995188509872305,0.022044752748883865,7,0
-Invocation,LightInvoke,Switch,1.510968037136868,0.01995102199199213,0.10789278304520543,8,0
-Invocation,SystemInvoke,CapturedConstant,1.0419648666720536,0.0032910366958241537,0.017797539813192806,8,0
-Invocation,FastInvoke,CapturedConstant,1.0429572324938534,0.0019130115577829154,0.010345341759920204,8,0
-Invocation,LightInvoke,CapturedConstant,1.0311589844091018,0.002251910363243133,0.01341871988491064,7,0
+phase,method,workload,mean_ns,standard_error_ns,confidence999_halfwidth_ns,samples,allocated_bytes
+Compilation,SystemCompile,Arithmetic,21057.400173611113,480.06713333032013,2596.146657766782,8,4447
+Compilation,FastCompile,Arithmetic,2136.472993116752,29.83925431808885,161.36718177467185,8,943
+Compilation,LightCompile,Arithmetic,1861.2931854812869,14.306173933474842,85.2478605594403,7,943
+Compilation,SystemCompile,Array,46886.166581483645,561.0764982423005,3034.2357860623824,8,5120
+Compilation,FastCompile,Array,2750.989670356446,24.45831453192665,145.74260011322139,7,1151
+Compilation,LightCompile,Array,2768.623869131542,29.66190207535884,160.40808168169767,8,1151
+Compilation,SystemCompile,Switch,29992.961711711712,233.24966158584692,1261.3867672017197,8,6182
+Compilation,FastCompile,Switch,2425.097851613464,42.57471854073646,230.23907610197057,8,1343
+Compilation,LightCompile,Switch,2286.3197476641903,10.490214069947784,56.729845278147806,8,1360
+Compilation,SystemCompile,CapturedConstant,34270.82298108068,315.7630793153843,1707.60963643542,8,4904
+Compilation,FastCompile,CapturedConstant,2065.5865263146625,23.705268182687245,128.19530538784718,8,1183
+Compilation,LightCompile,CapturedConstant,1975.462104590091,44.4838550235361,240.56346191163306,8,1183
+Construction,SystemTree,Arithmetic,440.8064759139369,1.9679978125448743,10.6427009657735,8,584
+Construction,LightTree,Arithmetic,178.25360727976607,0.9876112338033758,5.340885525790328,8,360
+Construction,SystemTree,Array,1058.828216831608,18.78200639040111,111.91852337604925,7,1192
+Construction,LightTree,Array,401.1927925192827,3.773679071513119,20.40761307909032,8,672
+Construction,SystemTree,Switch,1126.56504952069,10.951442884606463,65.25763494089202,7,1432
+Construction,LightTree,Switch,209.80873104222894,1.3395839536704643,7.2443126443456105,8,824
+Construction,SystemTree,CapturedConstant,508.16101190874906,4.123871927411091,22.301414902402737,8,528
+Construction,LightTree,CapturedConstant,237.72549601577592,0.8994772171331511,4.864267117805024,8,272
+Invocation,SystemInvoke,Arithmetic,1.1857331990510527,0.00393391632142369,0.021274157301601867,8,0
+Invocation,FastInvoke,Arithmetic,1.0989445437726775,0.00231112943902114,0.012498316489941222,8,0
+Invocation,LightInvoke,Arithmetic,1.0980764832439314,0.002266402297427581,0.012256437362840716,8,0
+Invocation,SystemInvoke,Array,1.0417457639223606,0.0023781178586783683,0.012860581993508882,8,0
+Invocation,FastInvoke,Array,1.2857391691880478,0.0006201253789979213,0.0036952131355322184,7,0
+Invocation,LightInvoke,Array,1.2984493339059076,0.004658099760368262,0.025190456260840082,8,0
+Invocation,SystemInvoke,Switch,1.0546706028009836,0.0016552381833178306,0.008951333634564273,8,0
+Invocation,FastInvoke,Switch,1.3091380151624001,0.0036995188509872305,0.022044752748883865,7,0
+Invocation,LightInvoke,Switch,1.510968037136868,0.01995102199199213,0.10789278304520543,8,0
+Invocation,SystemInvoke,CapturedConstant,1.0419648666720536,0.0032910366958241537,0.017797539813192806,8,0
+Invocation,FastInvoke,CapturedConstant,1.0429572324938534,0.0019130115577829154,0.010345341759920204,8,0
+Invocation,LightInvoke,CapturedConstant,1.0311589844091018,0.002251910363243133,0.01341871988491064,7,0
@@ -0,0 +1,33 @@
+phase,method,workload,mean_ns,standard_error_ns,confidence999_halfwidth_ns,samples,allocated_bytes
+Compilation,SystemCompile,Arithmetic,15844.082357296715,117.14190604498756,698.0270838748723,7,4447
+Compilation,FastCompile,Arithmetic,1810.91773726065,30.040842258692937,179.0078565985373,7,944
+Compilation,LightCompile,Arithmetic,1591.1956593496586,29.792845279752683,161.11620714758004,8,944
+Compilation,SystemCompile,Array,42645.951429263565,175.57477483085358,949.4877553961624,8,5120
+Compilation,FastCompile,Array,2393.7356108293993,52.03943593546823,310.09343230745947,7,1151
+Compilation,LightCompile,Array,2185.8164648564903,37.15748954054916,221.4146494947734,7,1152
+Compilation,SystemCompile,Switch,27327.010866195436,202.3703170054131,1094.39489948025,8,6182
+Compilation,FastCompile,Switch,2286.794093559805,45.01491264595569,268.2355893464061,7,1343
+Compilation,LightCompile,Switch,2129.9730526867133,45.06550584541042,243.70896122278313,8,1360
+Compilation,SystemCompile,CapturedConstant,33563.45956486043,367.30287574658945,2188.6903151275565,7,4903
+Compilation,FastCompile,CapturedConstant,1817.31042777688,9.800219336368553,58.397705446864656,7,1183
+Compilation,LightCompile,CapturedConstant,1757.592018363402,9.947656127690006,53.79575566725131,8,1183
+Construction,SystemTree,Arithmetic,446.7495686599703,2.8599120400153017,17.04169010758502,7,584
+Construction,LightTree,Arithmetic,176.39085490854032,0.8008337624067482,4.772021172565207,7,360
+Construction,SystemTree,Array,996.792736595607,3.130061731290863,16.927006116669833,8,1192
+Construction,LightTree,Array,404.4762214847132,2.945943735290582,17.554337164486697,7,672
+Construction,SystemTree,Switch,1126.901258079409,31.6513087179164,171.16655908502597,8,1432
+Construction,LightTree,Switch,223.43289142023625,5.741280784497018,31.048171984972665,8,824
+Construction,SystemTree,CapturedConstant,539.7029327203401,14.629401962529194,79.11408712084024,8,528
+Construction,LightTree,CapturedConstant,242.382570351469,2.0556169724482802,11.11653508885029,8,272
+Invocation,SystemInvoke,Arithmetic,1.1907134338045033,0.0029981832636031018,0.016213822856770626,8,0
+Invocation,FastInvoke,Arithmetic,1.1009212499043197,0.0007561604413512465,0.0045058210647113866,7,0
+Invocation,LightInvoke,Arithmetic,1.1085954129484215,0.0018091742396718748,0.009783801742599426,8,0
+Invocation,SystemInvoke,Array,1.0707570448998405,0.0035985578343859822,0.021443144610664336,7,0
+Invocation,FastInvoke,Array,1.3038193604005623,0.0023612641365360586,0.014070338917411785,7,0
+Invocation,LightInvoke,Array,1.3035723313280687,0.002988970637560874,0.016164002057452692,8,0
+Invocation,SystemInvoke,Switch,1.055336958921157,0.0026761006616824456,0.014472037984516533,8,0
+Invocation,FastInvoke,Switch,1.3488832816288703,0.01258963691991997,0.06808327740624193,8,0
+Invocation,LightInvoke,Switch,1.4580039992194078,0.010160983525320704,0.05494940517149252,8,0
+Invocation,SystemInvoke,CapturedConstant,1.0591227805490677,0.005009427756522489,0.029850259115848318,7,0
+Invocation,FastInvoke,CapturedConstant,1.0598954502524944,0.003608426486662622,0.02150195009542222,7,0
+Invocation,LightInvoke,CapturedConstant,1.0776547983562088,0.007377067469176524,0.03989431420010333,8,0
@@ -1,3 +1,7 @@
+# Historical screening results (original stack)
+
+These measurements belong to the original revisions below. For current three-stage FEC results after rebasing onto #558, see [Rebase-20261024](Rebase-20261024/README.md). System rows in these historical runs are unchanged reference controls; their variation does not mean System was modified. The original array optimization is already present in #558 and is no longer a #559 production change.
+
 # Local performance work, 2026-10-18
 
 Base: compatibility fixes `3f682768d6b23e445ae46595e456658899e371e4`.
 
@@ -0,0 +1 @@
+* -text
@@ -0,0 +1,78 @@
+# Three-stage comparison after rebasing #559 onto #558
+
+## Exact revisions and retained scope
+
+- Before #558: current upstream master without #558, `233ed5adaa9d9ceea67d71e153d0d2457f2cce00`. This includes merged #557/#562 and upstream SmallMap/conversion-helper repairs. It is not the historical original parent of #558; choosing this matched base isolates the current PR's incremental cost.
+- After #558: `aaa79fe73f0978e9be8150116d4175f04cb5c971`.
+- After #559: rebased production/test source `e300f4e2`.
+
+Rebased the performance layer from `3f682768` onto current #558. The array-classification commit `7b6863bc` was dropped: #558 already uses IsSZArray on modern configured targets and retains its Framework fallback. The remaining production change is `d4a74798`, named LightExpression field lookup using GetDeclaredField with existing special-metadata handling. Native runner conflicts retain all #558 regressions and add the named-field/conditional oracles. No #558 branch rewrite or source change was made. The earlier conditional branch candidate remains rejected.
+
+Historical measurement CSVs are preserved; dropped source does not erase historical evidence. The old Summarize confidence-column correction is retained. Historical System rows are unchanged reference controls, not changes to System caused by this PR. Current results below supersede historical timing claims for this rebased source.
+
+## Compilation comparison
+
+Ratios above 1 mean slower. Each value is the descriptive median of 41 workload-specific time ratios, not an aggregate application result. The ratio columns have independently computed medians, so they do not multiply exactly.
+
+| Runtime/flavor | Before #558 | After #558 / before | After #559 / before | After #559 / #558 |
+| --- | ---: | ---: | ---: | ---: |
+| 8.0.28 standard | 1.000x | 1.0461x | 1.0890x | 1.0395x |
+| 8.0.28 light | 1.000x | 1.1690x | 1.2031x | 1.0441x |
+| 10.0.9 standard | 1.000x | 1.1164x | 1.2042x | 1.0375x |
+| 10.0.9 light | 1.000x | 1.0746x | 1.1041x | 1.0303x |
+
+#558 still has measured compilation costs relative to current master: these medians are about 4.6–16.9% higher. #559 does not establish a compilation improvement: its medians are 3.0–4.4% higher than #558 in this run. All Standard production source files are byte-identical between #558/#559, so the Standard timing movement cannot be attributed to a Standard compiler code change. Light's compiler/emitter source is also identical; its expression factory implementation differs. These observations retain run-to-run/JIT/layout/host variation and do not establish a causal CPU regression or universal parity. See every workload and raw distribution rather than treating a median ratio as a guarantee.
+
+Adaptive operation counts produced small fractional allocation residuals. A separate equal-count check (200 compilations/sample, 100 warmups, seven samples in each of two opposite-order rounds) confirms for **each runtime/flavor**: #558 allocates less in **14/41** workloads and the same in **27/41**, with **zero increases**; #559 matches #558 in **all 41**. Original adaptive residuals remain in the files rather than being rounded away. `allocation-confirmation.json` contains the equal-count values.
+
+## Remaining optimization: Light tree construction
+
+The named-field change affects construction, not a text parser. The broad construction pass caps at 20,000 operations; short fast samples were therefore confirmed separately with a higher cap (2,000,000) and a 30 ms calibration target. Both datasets remain available. Longer-run medians for the three selected affected shapes follow; unchanged constant/array and additional capture controls are retained in `construction-confirmation.json`.
+
+| Light construction / runtime | After #558, ns | After #559, ns | B/op after #558 → #559 |
+| --- | ---: | ---: | ---: |
+| new_member / 8.0.28 | 127.0 | 108.2 | 304.01 → 272.01 |
+| capture_reference / 8.0.28 | 139.2 | 97.3 | 368.01 → 336.01 |
+| capture_mutable_reference_ctor / 8.0.28 | 226.6 | 178.3 | 704.03 → 672.03 |
+| new_member / 10.0.9 | 116.5 | 109.6 | 304.01 → 272.01 |
+| capture_reference / 10.0.9 | 97.5 | 92.1 | 368.02 → 336.01 |
+| capture_mutable_reference_ctor / 10.0.9 | 188.5 | 171.4 | 704.03 → 672.03 |
+
+These affected workloads save approximately **32 B per lookup** on this host/runtime, rather than reusing the old report's 40 B figure. Observed time reductions are roughly 5.6–30.1% across the six entries; unchanged controls also move, so this is not a universal percentage claim. Other workload construction values, including slower/noisy ones, remain in the full data.
+
+## Generated delegate execution
+
+After #558 vs after #559, all **82 root delegate IL bodies** match for 41 workloads in both flavors on .NET 10. This checks root IL, not every nested delegate body or JIT-generated machine code. The compiler source is unchanged. No delegate runtime speedup is claimed. Broad runtime medians (after559/after558) range about 0.994–1.039x; the raw samples and each workload's allocation results are in the phase files. For factory workloads, `invoke` includes outer activation plus invocation of the resulting inner delegate, rather than only a scalar call. All generated delegates pass an untimed System behavior oracle for the tested inputs.
+
+## System: separate unchanged reference
+
+System was previously measured in multiple process runs as a timing-drift control. Those varying numbers did not represent patches to System; presenting them as optimization before/after columns was misleading. Here the main comparison has exactly three FEC revisions. System appears only as **one separately measured unchanged reference per runtime**, pooled over two reference process repetitions. It has no before/after delta, and no optimization claim is normalized to it. The standalone reference executable uses only System, without a FEC library dependency. There are no System construction rows in this reference. Sample compilation references:
+
+| Unchanged System reference | Workload | Compile us |
+| --- | --- | ---: |
+| 8.0.28 | new_member | 17.118 |
+| 8.0.28 | array | 30.947 |
+| 8.0.28 | capture8 | 194.630 |
+| 8.0.28 | balanced256 | 246.676 |
+| 10.0.9 | new_member | 14.402 |
+| 10.0.9 | array | 29.845 |
+| 10.0.9 | capture8 | 154.196 |
+| 10.0.9 | balanced256 | 220.502 |
+
+## Method and validation
+
+Windows ARM64, .NET 8.0.28 / 10.0.9, SDK 10.0.301, Release, tiering and ReadyToRun disabled. Separate frozen source/binary snapshots, serial processes, forward/reverse FEC revision order, seven samples per phase per round. 41 common workloads cover scalars, arrays, switches, blocks, scopes/EH, invocation and captures. Construction, compilation of a prebuilt tree (including delegate creation), and warmed execution are separate. Light trees are constructed directly. Allocation uses the current thread's managed allocated-byte counter. There are 20,664 broad FEC observations plus 2,296 System-reference observations; 6,888 equal-count compilation observations and 1,176 longer construction observations are retained separately. Six non-common/new-feature benchmark shapes are listed in source-manifest.json and were not timed. They do not change the imported test exclusions. No fresh cold/first-invocation, Framework or x64 timing, retained-memory or native/JIT allocation claim is made.
+
+Validated source `e300f4e2`: all production targets build; full .NET 10/.NET 8 ARM64 native suites pass (1,714 / 1,713 legacy cases plus registered regressions); Framework passes 1,049 TestX + 1,597 legacy cases. All 21 imported batches pass: **35,064 executed passes, zero executed failures/crashes, 69 existing exclusions unchanged**. The deferred deep-tree modes remain deferred. The retained BenchmarkDotNet project builds. The evidence-only commit changes no validated production/test code.
+
+## Files
+
+- `three-stage.csv`: absolute timings, allocations and three-stage ratios for every workload/phase/configuration.
+- `phase-medians.csv`: all FEC phase medians and the separate System references, including ranges/sample operation counts.
+- `allocation-confirmation.json`: equal-count compilation allocation proof.
+- `construction-confirmation.json`: longer construction controls.
+- `analysis.json`, `summary.csv`: full analysis; adaptive allocation counts are screening observations, superseded by equal-count confirmation for allocation conclusions.
+- `validation.json`: revisions, counters and source/assembly hashes; `il-comparison.json`: root delegate comparisons.
+- `source-manifest.json`, `binary-manifest.json`, `system-reference-manifest.json`: immutable snapshot identities.
+- `range-diff.txt`: old/rebased performance-layer mapping.
+- `evidence.zip` and `evidence-manifest.json`: complete sources, harnesses/scripts, project files, native/imported validation logs, raw measurements, confirmation data and manifests. Binaries are hashed and retained locally, not committed. Runners refuse to mix existing output with a new run; preserve/relocate archived output before replaying in a fresh directory. The initial System-reference build error is retained; the corrected reference project was then built and measured successfully.